# A Self-Improving Coding Agent
**Authors**:
- Maxime Robeyns (University of Bristol)
- &Martin Szummer
- iGent AI
- Laurence aitchison (University of Bristol)
## Abstract
Recent advancements in Large Language Models (LLMs) have spurred interest in deploying LLM agents to undertake tasks in the world. LLMs are often deployed in agent systems: code that orchestrates LLM calls and provides them with tools. We demonstrate that an agent system, equipped with basic coding tools, can autonomously edit itself, and thereby improve its performance on benchmark tasks. We find performance gains from 17% to 53% on a random subset of SWE Bench Verified, with additional performance gains on LiveCodeBench, as well as synthetically generated agent benchmarks. Our work represents an advancement in the automated and open-ended design of agentic systems, and demonstrates a data-efficient, non gradient-based learning mechanism driven by LLM reflection and code updates.
## 1 Introduction
LLM s have recently made impressive advancements across a range of domains and tasks [2, 14, 29]. However, in order to put these LLM s to use in real world applications, LLM s must be wrapped in code to orchestrate them and expose tools that allow the models to take actions. These action-taking LLM s are referred to as agents, and the broader system an agent system.
These agent systems often show dramatic improvements in benchmark performance over âplainâ LLM s [43, 46, 5], through combinations of prompting strategies and methods for combining different LLM outputs. Early examples include best-of- $N$ sampling and simple prompting strategies such as chain of thought [20]. However more sophisticated schemes have shown success in getting the desired behavior and performance improvements from the models, for instance STaR [45], Tree of Thoughts [42], Graph of Thoughts [3], LLM Debate [10], Iterative Self-Refinement [23], Expert Prompting [22] among many others. The comprehensive survey of Schulhoff et al. [34] demonstrates the vast number of manually created strategies to date.
Recent improvements in coding agents raise the question of whether these agents themselves can autonomously modify and improve their own code by discovering e.g. new prompting schemes or tools without manual design and implementation. We argue that this style of fully self-referential meta-agent programming is possible today and offers a sound alternative to the ad-hoc, trial-and-error approach of hand-crafted orchestrators, which may only explore a small fraction of the solution space. Recent work in the Automated Design of Agentic Systems (ADAS) [16] uses a meta-agent to optimise agent implementations. However, Hu et al. [16] is not self -improving, as there are two separate agents: the target-agent that performs the task, and the meta-agent, which improves the target agent. A motivation for a self-improving system is that the improvements in coding abilities may be leveraged during subsequent improvement steps, hopefully compounding.
Our contributions are:
- A self-improving coding agent (SICA) that eliminates the distinction between meta-agent and target agent, and is capable of editing its own codebase to improve itself with respect to its cost, speed and benchmark performance.
- Empirical evidence that self-referential agents can effectively improve their own implementations; we find performance improves from 17% to 53% performance on a random subset of SWE-Bench Verified, even with consideration given to safety constraints and resource efficiency.
- We share a our implementation of a self-improving coding agent (SICA) with the community. SICA is implemented in standard Python without a domain-specific language, and provides a reference agent framework for building new SICA systems, as well as those seeking to post-train LLMs on tool use and other agent tasks.
We make our code available at https://github.com/MaximeRobeyns/self_improving_coding_agent.
## 2 Related Work
<details>
<summary>x1.png Details</summary>
### Visual Description
## Flowchart: Agent Development and Benchmarking Process
### Overview
The diagram illustrates a multi-agent iterative improvement process where each agent's code is refined through benchmark evaluations and meta-improvement cycles. Three agents (0, 1, 2) are shown with explicit connections, suggesting a scalable system (indicated by "..." for additional agents).
### Components/Axes
1. **Agents**:
- Agent 0 (Base Agent)
- Agent 1
- Agent 2
- ... (implied continuation)
2. **Code Components**:
- Base Code (Agent 0)
- Agent 1 Code
- Agent 2 Code
3. **Evaluation Framework**:
- Benchmarks (3 per agent: Bench 1, Bench 2, Bench 3)
4. **Meta-Improvement Arrows**:
- Directed from Agent 0 â Agent 1 â Agent 2 â ...
- Label: "Meta-Improvement"
### Detailed Analysis
- **Agent 0**:
- Contains "Base Code" (highlighted)
- Three benchmarks (Bench 1-3)
- Labeled as "Base Agent"
- **Agent 1**:
- Contains "Agent 1 Code"
- Three benchmarks (Bench 1-3)
- Connected to Agent 0 via "Meta-Improvement" arrow
- **Agent 2**:
- Contains "Agent 2 Code"
- Three benchmarks (Bench 1-3)
- Connected to Agent 1 via "Meta-Improvement" arrow
- **Best Agent Selection**:
- "Best Agent 0, 1" (bottom-left)
- "Best Agent 0, ..., 2" (bottom-right)
- Indicates iterative selection across agents
### Key Observations
1. **Iterative Refinement**: Each agent's code is positioned as an improvement over the previous through meta-improvement cycles.
2. **Benchmark Consistency**: All agents share identical benchmark structures (3 per agent), suggesting standardized evaluation criteria.
3. **Scalability**: The "..." notation implies the system can accommodate additional agents beyond Agent 2.
4. **Hierarchical Selection**: The "Best Agent" labels indicate a comparative evaluation process across agents.
### Interpretation
This diagram represents an evolutionary optimization framework where:
1. **Base Agent (Agent 0)** serves as the initial reference point
2. **Meta-Improvement** arrows suggest knowledge transfer or algorithmic refinement between agents
3. **Benchmarking** acts as the evaluation mechanism for code performance
4. **Best Agent Selection** implies a competitive process where agents are ranked based on benchmark results
The structure suggests a reinforcement learning or genetic algorithm approach where:
- Each new agent incorporates improvements from previous iterations
- Benchmarks provide objective performance metrics
- The "Best Agent" selection represents the global optimum across iterations
Notably, the absence of quantitative performance metrics in the diagram leaves the exact nature of "improvement" undefined - it could represent speed, accuracy, resource efficiency, or other measurable criteria. The consistent benchmark structure across agents implies standardized evaluation parameters, while the meta-improvement arrows suggest cumulative knowledge transfer between iterations.
</details>
Figure 1: Meta Agent Loop: the agents starts with the minimal code required to support initial self-improvement, and then follows a sequence of benchmarking and meta-improvement.
The traditional approach to developing and optimizing agent systems has been to manually design agent architectures and prompting techniques. Notable examples include Chain-of-Thought prompting [39], self-refinement [23] and self-reflection [37] for improving reasoning, tool use frameworks [33], and various compositional agent systems [38, 1]. While these hand-crafted approaches have achieved strong results, they require significant human effort and may miss useful patterns that could be discovered through automated search.
Another direction focuses on enabling agents to learn reusable skills and continuously self-improve. MaestroMotif [19] uses LLM feedback to learn skill rewards and combines skills through code generation. This builds on earlier work on intrinsically motivated reinforcement learning [6] and autotelic agents [8] that develop repertoires of internally motivated skills, as well as work in open-endedness [48, 12] that use LLM s to identify interesting and useful directions to explore in.
Several approaches leverage LLMs to optimize agent behaviors through natural language interaction. OPRO [41] and Promptbreeder [13] focus on optimizing prompts through language. Others have explored using LLMs to critique and refine agent behaviors [18], generate training curricula [21], or provide natural language feedback for reinforcement learning [31]. Since our agent can edit its entire codebase, this includes the ability to tune its own prompts to optimise the behaviour of any part of the agent.
Recent work has begun exploring automated approaches for designing and optimizing agent systems. AgentSquare [35] proposed a modular design space that abstracts agent components into planning, reasoning, tool use and memory modules, allowing automated search over module combinations. AlphaEvolve [28] set out to use a coding agent to make scientific discoveries and optimize computational infrastructure, the results of which may accelerate the training of the LLM underpinning AlphaEvolve itself. The approach taken by these works leans more heavily on structured, evolutionary search than our own.
Perhaps the most closely related line of prior work began with ADAS [16]. ADAS used a target-agent which performs the actual task, and a meta-agent which improves the target-agent. As such, ADAS is not self-improving (as the meta-agent improves the target agent, not itself). Moreover, in ADAS the meta-agent edits only a single forward function, written in a domain-specific language which has been carefully designed to make expressing different prompting schemes very straightforward. In contrast, our self-improving coding agent is fully self-improving (i.e. there is no distinction between the meta and target agent), and it operates over the agentâs full Python codebase.
Of course, we would expect the first truly self-improving agents to be coding agents, because agents are written in code. The natural approach is to start off with a basic coding agent that can open/close/edit files, run commands in the terminal etc, then to launch this agent in a self-improvement loop. We believe that our self-improving coding agent is the first such work. However, there are two papers claiming self-improving agents, but they do not evaluate in the coding setting, as they do not consider âfullâ coding agents. First, GĂśdel Agent [44] has specific tools (such as action_adjust_logic and action_read_logic) that allow modification of small parts of the agent as it is running. Thus, it is not a general-purpose coding agent, as it is traditionally understood. And as such, as with ADAS, it was evaluated on language understanding and mathematical benchmarks (DROP [11], MGSM [36], MMLU [15] and GPQA [32]), rather than coding benchmarks. Second, Zelikman et al. [47] introduce a self-taught optimizer for recursively self-improving code generation. Again, this is not a general-purpose coding agent with the ability to e.g. work on arbitrary software engineering tasks, and as such, it self-improves and evaluates and only on algorithmic tasks such as learning parity with noise [4], String Grid Distance, and 3-SAT.
## 3 Methods
The main running loop of SICA resembles the ADAS loop Hu et al. [16] and is illustrated in Figure 1. In particular, both SICA, and ADAS keep an archive of previous agents and their benchmark results. In the first step, SICA takes the best performing agent from the archive up until this point as the meta-agent, which is instructed to look through the archive, identify an improvement, and implement it. The high-level scheme is illustrated in Algorithm 1. Note that this is different from ADAS, in that ADAS has a fixed meta-agent, so does not select the meta-agent from the archive (the archive in ADAS is an archive of target-agents, not meta-agents). Second, both ADAS and SICA evaluate the new agent on a set of evaluation benchmarks and store the results.
We define the âbestâ performing agent by a utility function, which incorporates common desiderata for a good agent; the normalised benchmark performance score $p_{\text{score}}\in[0,1]$ , the wall-clock time in seconds $p_{\text{time}}$ , as well as the dollar cost $p_{\text{cost}}$ . The base utility is calculated as
$$
U=w_{\text{score}}p_{\text{score}}+w_{\text{cost}}(1-\min(1,p_{\text{cost}}/\mathdollar 10))+w_{\text{time}}(1-\min(1,p_{\text{time}}/300\text{ s})) \tag{1}
$$
where we set the coefficients to $w_{\text{score}}=0.5$ , $w_{\text{cost}}=0.25$ and $w_{\text{time}}=0.25$ . We cap the worst cost at $10 per problem, and impose a timeout of 300 s after which we cancel the agent. To give partial credit for work done before the timeout, we calculate the final utility as follows, with a timeout penalty of $\tau=0.5$ :
$$
U_{\text{final}}=\begin{cases}U(1-\tau),&\text{if timed out}\\
U,&\text{otherwise}\end{cases} \tag{2}
$$
Note that since we are not performing any weight updates, this numerical score is only used to pick the next meta agent, as well as the base agent for the next iteration.
We begin by describing the initial coding agent. We then go on to describe the benchmark runner framework, and how this framework naturally allows us to create a self-referential task (that of improving the coding agent). System Prompt
⢠Agent definition and goals ⢠Tool documentation ⢠Sub-agent documentation ⢠System instructions Core Prompt
⢠Problem statement or instruction ⢠Open files ⢠Directory tree Assistant Messages
⢠CoT reasoning ⢠Tool calls and results ⢠Sub-agent calls and results ⢠Overseer notifications LLM Context Window
Figure 2: LLM context window structure.
### 3.1 Initial agent
We were interested in useful, practically relevant coding agents capable of self-improvement. To that end, we started with a sophisticated coding agent, with a number of tools, sub-agents and oversight mechanisms. Of course, the tools, sub-agents and oversight mechanisms were all amenable to self-improvement; but here we simply describe the starting point.
We designed a set of tools to open (i.e. add to the context) and close files, overwrite files, execute a shell command and a simple calculator tool. We also include basic tools to submit a final answer to a benchmark problem, return a result from a sub-agent call to its caller, as well as exiting early from a sub-agent call. Finally, we provide an âarchive analysisâ tool, which facilitates extracting high-level summary statistics about each agent iteration in the archive. We describe the tool calling mechanism in more detail in Appendix C.
Next, we included several sub-agents. These are merely agents invoked by an existing agent, much like a tool call, by accepting input arguments and returning a string value. Introducing this abstraction serves two purposes. The first is to ensure the benchmark tasks can run entirely autonomously, without human direction and guidance. We provide the problem statement to the main entrypoint agent, (the full prompt for which is given in Appendix A) which then decomposes the task, delegating and asking questions to sub-agents. The second is to manage the context length for extended agent runs. The meta-improvement task is challenging, and the main agent may take on the order of one hour to finish one meta-improvement iteration. The sub-agent calls allow the context used for a sub-task to be naturally discarded when the agent returns, with the value returned by the sub-agent providing a summary of that sub-task in the main agentâs context. Initially, we included three agents: a coding agent, a problem-solver agent and a reasoning agent. The full prompts for these agents are given in Appendix A.1.
Finally, one important component we provide in the initial agent that enables extended, autonomous operation while remaining on-task, avoiding loops or other behaviours that frustrate progress is the asynchronous overseer. This is an LLM, running periodically in a concurrent thread, which is prompted to judge whether there are any pathological behaviours or deviations from the original task, and which may intervene by notifying the currently running agent (or any of its parents) by means of a message in-lined into the assistant context, or in serious cases to cancel the execution of an agent if it has deviated too far or appears stuck. The overseer is called every 30s by default, or on key events.
This overseer is provided with a comprehensive textual representation of the state of the system (a truncated example of which is provided in Appendix B) with verbose content truncated to limit context. This consists of the âcallgraphâ of agent and sub-agent execution, in addition to the event stream of LLM messages, tool calls and responses, and sub-agent calls and responses associated with each agent call. This broad representation of the systemâs behaviour, allows the overseer to to pick out patterns, as well as detect repeated work completed by previous agent calls. The full Overseer prompt is given in Appendix A.2.
The structure of the agentâs contexts, which contains e.g. open file contents, rather than just the prompt is critical. In the initial coding agent, the context is structured as Figure 2. The system prompt with the agent definition (from Appendix A) is presented first, listing the definitions of the tools that the agent has available and the sub-agents it has available to call too. The system prompt concludes with system information, such as instructions for how to break out of the agent loop and return to the calling process.
Next comes the âcore promptâ, which is set as the first user message in the chat template format We use the user message role in the chat template for the core prompt not by design, but owing to the fact that we are using chat endpoints from model provider APIs. If fine-tuning the LLM specifically for agentic applications, one might choose a more appropriately named section delimiter such as <core>., and contains the caller-specified problem statement to work on (the caller might be the user invoking the agent, or an agent calling a sub-agent). This is also where we insert a view of the files the agent has open, as well as a representation of the current working directory.
Finally, the remainder of the LLM agentâs context is the running assistant message. This contains periods of chain-of-thought reasoning about the problem, punctuated by tool calls and their responses, sub-agent calls and their responses, as well as any notifications or comments from the asynchronous overseer system.
The LLM context is laid out in this way to allow new content to be appended, hence preserving the KV cache to the benefit of latency between tool calls and cost. For the base agent, we adopt a simple scheme where file edits are appended as diffs, and these are periodically re-consolidated into the stale file view after a certain number accumulate. More sophisticated KV cache preservation schemes may be written by the agent to drive down cost.
Input: Evaluation benchmarks $B=\{b_{j}:j=1,\ldots,m\}$ , iteration count $n\in\mathbb{N}$
Output: Improved agent system $A_{n}$
1 Initialise agent $A_{0}$
2 for $i=0,1,\ldots,n-1$ do
3 Evaluate $A_{i}$ on benchmarks $B$ and store in $p_{i}$
4 Run $A_{\hat{i}}$ to generate $A_{i+1}$ , using the archive of $A_{\leq i}$ and $p_{\leq i}$ , where $\hat{i}=\operatorname*{arg\,max}_{j\in\{0,\ldots,i\}}p_{j}$
5
6 return $A_{n}$
Algorithm 1 Self-Referential Agent Improvement.
## 4 Experiments and Results
$0$ $1$ $2$ $3$ $4$ $5$ $6$ $7$ $8$ $9$ $10$ $11$ $12$ $13$ $14$ $15$ $0.52$ $0.54$ $0.56$ $0.58$ $0.6$ $0.62$ $0.64$ $0.66$ $0.68$ $0.7$ âSmart Editâ Tool Code Context Summarization File Edit Verification AST Symbol Locator Hybrid Symbol Locator Agent Iteration Utility Mean Utility Best Utility So Far 95% Confidence Interval
Figure 3: Performance across iterations. Key improvements are annotated with their corresponding tool or agent modifications.
We first show a run that includes relatively long-horizon tasks in the benchmark set. We include SWE Bench Verified [7] which tests problem decomposition, code navigation as well as fast and efficient file editing from which we select a fixed random subset of 50 questions. We also include 50 random questions from LiveCodeBench [26], the questions from which resemble competitive programming and often include more theoretical reasoning Both SWEBench-Verified and LiveCodeBench are MIT licensed.
We also include two example âsyntheticâ benchmark tasks, defined and curated entirely within the agentâs codebase. The first of these is a file editing benchmark, constructed by cloning the repositories used in SWE bench, finding âinterestingâ (i.e. not trivial, nor complete overwrites) file edits in consecutive commits, mounting the repository checked out at the first commit in the working directory, and prompting the agent to edit the file such that it matches the target content of the second commit. The agent is scored on the closeness of the final file content to the target content, with time and dollar cost and token consumption also being accounted for in the utility function. The second is a codebase navigation problem, which is also constructed by cloning some Python repositories, from which we identify âinterestingâ symbols, and at least one reference to this symbol in the project. The benchmark task is for the agent to locate the definition of this symbol, and return the location in the format path/to/file.py:line_num:column_num from the example reference.
We run the agent system using Sonnet 3.5 (v2) [2] in most of the agents in the agent system, with the exception of a âreasoning agentâ which uses o3 mini [30]. While the base agent system can run with much smaller and less expensive models, and many of our benchmark tasks can be solved by this class of models, the meta-improvement task is a complicated long-horizon task which requires powerful models. The API cost of the 15 iteration run was approximately $7,000.
The results are shown in Figure 3, with more detailed iteration-level peformance numbers in Table 1.
Table 1: Performance per iteration: the benchmark accuracy numbers correspond to a synthetic file editing benchmark and symbol location (codebase navigation) benchmark, SWE-Bench Verified, and LiveCodeBench, respectively. We also show the metrics averaged across all problem instances across all benchmarks to compare the relative âper problemâ costs across agent iterations.
| | Benchmark Accuracy | Average Metrics (per benchmark problem) | | | | | | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Iter | File | Sym | SWE-Bv | LCB | Cost ($) | Time (s) | Tokens (M) | % Cached |
| 0 | $0.82$ | $0.35$ | $0.17$ | $0.65$ | $1.91$ | $130.2$ | 0.24 | 32.5 |
| 1 | $0.87$ | $0.32$ | $0.14$ | $0.62$ | $1.71$ | $123.8$ | 0.24 | 33.6 |
| 2 | $0.92$ | $0.31$ | $0.17$ | $0.58$ | $2.45$ | $151.4$ | 0.26 | 34.3 |
| 3 | $0.82$ | $0.33$ | $0.22$ | $0.64$ | $1.84$ | $126.9$ | 0.29 | 31.9 |
| 4 | $0.88$ | $0.31$ | $0.38$ | $0.54$ | $2.70$ | $148.3$ | 0.26 | 30.9 |
| 5 | $0.89$ | $0.31$ | $0.30$ | $0.59$ | $2.17$ | $134.8$ | 0.23 | 36.9 |
| 6 | $0.96$ | $0.31$ | $0.37$ | $0.64$ | $2.21$ | $134.1$ | 0.25 | 36.1 |
| 7 | $0.92$ | $0.35$ | $0.33$ | $0.58$ | $2.15$ | $134.9$ | 0.23 | 38.2 |
| 8 | $0.93$ | $0.33$ | $0.27$ | $0.64$ | $2.03$ | $128.5$ | 0.26 | 36.4 |
| 9 | $0.88$ | $0.40$ | $0.47$ | $0.61$ | $2.03$ | $126.3$ | 0.27 | 38.3 |
| 10 | $0.87$ | $0.41$ | $0.46$ | $0.66$ | $1.77$ | $107.0$ | 0.22 | 40.9 |
| 11 | $0.89$ | $0.43$ | $0.44$ | $0.70$ | $2.25$ | $129.6$ | 0.27 | 36.9 |
| 12 | $0.91$ | $0.38$ | $0.44$ | $0.64$ | $1.58$ | $103.9$ | 0.26 | 39.0 |
| 13 | $0.86$ | $0.40$ | $0.27$ | $0.61$ | $1.66$ | $113.3$ | 0.29 | 37.2 |
| 14 | $0.94$ | $0.34$ | $0.53$ | $0.67$ | $2.20$ | $117.1$ | 0.25 | 37.2 |
| 15 | $0.91$ | $0.40$ | $0.51$ | $0.71$ | $1.70$ | $114.5$ | 0.30 | 36.3 |
The numbers show a promising progression. For the file editing benchmark, the agent is initially only equipped with the ability to overwrite files which is slow and token inefficient, and it quickly implemented self-termed âsmartâ editing tools based on diffs and ranges. For the symbol navigation task, the agent implemented a simple AST based symbol locator at iteration 9 (perhaps exploiting the fact that our benchmarks were predominantly Python based), which yielded good improvements also reflected in other tasks. Note that we attribute the low score in the symbol locator benchmark to poor data qualityâsome of the target symbols point to uninstalled external libraries for instance. The SWE Bench Verified subset saw an appreciable increase in performance throughout the run, accuring agent framework improvements. The LiveCodeBench scores saw a subtle improvement, although not as pronounced as some of the other benchmarks.
### 4.1 Performance on Reasoning Tasks
We also evaluate the effectiveness of the self-referential agent system in improving task performance in more reasoning-heavy domains. We ran another experiment with two question answering tasks in the benchmark set: AIME 2024 and GPQA Diamond [32], the results of which are shown in Figure 4. $0$ $1$ $2$ $3$ $4$ $5$ $6$ $7$ $8$ $9$ $0.65$ $0.7$ $0.75$ $0.8$ Independent math verifier Sympy symbolic calculator Math cross validator Geometry specialist Systematic mathematical reasoning Agent Iteration Mean Accuracy AIME and GPQA Show Little Improvement Mean Accuracy Best Mean Accuracy So Far 95% Confidence Interval
Figure 4: Agent Framework Saturation: the benefits the agent system was able to find when the models alone (e.g. o3-mini-high) already perform well was marginal.
The results here show less improvement, and highlight an important interplay between the base models and the scaffolding system. The base agent system at iteration 0 used in Figure 4 used Sonnet 3.5, with a âreasoningâ sub-agent that used o3-mini. The o3-mini model alone scores 87% and 79% on AIME and GPQA Diamond with a âhighâ reasoning effort, while the agent system as a whole averaged 76% across the two benchmarks.
Inspecting the traces, we observe that for many of the runs, the main agent merely delegated the problem to the o3-mini based reasoning agent, and did not leverage the full capabilites of the agent system. For âreasoning modelsâ such as o1 [29], o3-mini or DeepSeekâs R1 [9]), we suspect that the inclusion of crude components aiming to induce reasoning behaviour (such as those included in iterations 4 and 6 in Figure 4) may in fact interrupt the reasoning chain of thought of a reasoning model trained outside the agent system, resulting in a drop in performance. We look forward to future work training âagentâ LLM s jointly with the design of the agent system which will hopefully make better use of both componentsâ strengths.
## 5 Conclusion
We provide SICA, a self-improving coding agent. Our results demonstrate that self-improvement, implemented via an agent system editing its own codebase, is indeed possible, particularly in longer-horizon âagenticâ tasks where the base LLM s stand to gain more from the structure and guidance of a scaffolding systemâfor instance showing improvements from 17% to 53% on a random subset of SWE-Bench Verified, while also achieving a slight decrease in the average time spent per problem on average. The framework we present offers both a practical approach to automated agent system design and a foundation for future work in jointly training foundation models with their agent systems.
### 5.1 Limitations
Our initial attempt at a self-improving coding agent is not without limitations. One key difficulty was getting the LLM-based agent to autonomously come up with truely novel, innovative, feasible and interesting modification ideas at each meta-improvement step, which is a theme which has been commented on in the open-ended learning literature [27, 40, 17]. The cost of settling on a bad idea which suffered from poor âtasteâ was a lengthy agent editing step followed by an even more expensive run through the benchmarks. While the failed iteration persists in the archive, in principle acting as an example of what not to do, we found that the initial feature ideas would often heavily influence later feature ideas as variations on the same theme. This path dependency may lead to higher variance agent runs; with poor quality initial feature suggestions (e.g. fixating on caching open files) often lowering the quality of subsequent feature suggestions.
We also note that in optimizing for agent running time and cost, our relatively short 5-minute timeouts (and to a lesser degree per-problem cost limits) cause the initial benchmark scores to perhaps be lower than expected for the underlying language model (e.g. Sonnet 3.5 v2), especially for longer-horizon benchmark tasks like SWE-Bench. Indeed, much of the initial improvement we report arises from features speeding up file editing and agent cost, which are not traditionally reflected to the same degree in SWE-Bench numbers.
### 5.2 Future Work
Our approach reflects a non weight-based learning mechanism, where we update the systemâs behaviour through a combination of optimising the prompts on which the model is conditioned, as well as the way the LLM calls are orchestrated by editing the underlying code. This is in some sense very sample-efficient, where a detailed inspection of the archive of outcomes informs substantial modifications at each step. However, we recognize the potential of jointly updating the foundation modelâs weights with the agent system. The work of AlphaEvolve [28] demonstrates a proof of this principle, and it may be fruitful to more explicitly fine-tune the foundation model to work with new newly created tools and scaffold at each iteration.
While we did not get to the point of benchmark saturation in our relatively short runs, we expect that future systems may run out of things to learn from a static benchmark set. One way this problem may be solved is through the automated design and curation of benchmarks, as initially demonstrated in the work of [12] where RL environments are designed in code. By including two synthetic benchmarks in our work which were curated entirely by the agent through manual prompting (the file editing and symbol navigation benchmark) we demonstrate that SICA is flexible enough to support creating new benchmarks and evaluations for itself.
## 6 Safety Considerations
It is vitally important to carefully consider the safety implications posed by systems designed to be self-improving. While our self-improving coding agent shows improvement in certain agentic tasks, there seem to be limits to what can be achieved with such purely scaffolding based systems (as seen in Figure 4), which do not modify the model weights. If the system updated the model weights as well as the scaffolding, we would expect that the risks of recursively self-improving systems would be higher. We highlight two key mechanisms for mitigating any risks.
First, a key safety mitigation in such systems is observability: that humans can and do oversee all aspects of potentially self-improving systems. We have thus architected SICA with sophisticated yet user-friendly observability mechanisms, which provide rich insight into the chain-of-thought, actions and sub-agent invocations at each step. These include an interactive web interface for easy oversight of all events occurring in the agent system, as well as the asynchronous LLM-based overseer, which may be prompted to treat various behaviors as grounds for canceling an agent, and which may also use a different LLM to the running agent. See Appendix A.2 for an example overseer judgement and the control afforded to it at each asynchronous invocation. Given the lack of model weight updates, which make it difficult but not impossible for the model to encode e.g. stenographic messages [24, 25] in the chain-of-thought, we believe this provides strong safety mitigation. That said, for recursively self-improving systems where the model weights are also updated, observability mechanisms, on their own, are clearly inadequate, and further research on safety mechanisms is clearly necessary.
The second opportunity for safety mitigation is inherent to the iterative, evaluation-based update procedure. While our initial benchmark set focuses on measuring capabilities we wish to improve, it is entirely possible to include safety-related evaluations in this benchmark set, to validate each agent before it progresses to the next iteration as the meta-agent.
We do not believe there to be any significant, direct adverse societal consequences to this work. Our objectives either focus on improving the mechanics of code editing or the effectiveness of multi-step reasoning through longer-horizon coding tasks.
## References
- Ahn et al. [2022] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil J. Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do As I Can, Not As I Say: Grounding Language in Robotic Affordances, August 2022. URL http://arxiv.org/abs/2204.01691.
- [2] Anthropic. Introducing Claude 3.5 Sonnet. URL https://www.anthropic.com/news/claude-3-5-sonnet.
- Besta et al. [2023] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Michal Podstawski, Hubert Niewiadomski, Piotr Nyczyk, and Torsten Hoefler. Graph of Thoughts: Solving Elaborate Problems with Large Language Models, August 2023. URL http://arxiv.org/abs/2308.09687.
- Blum et al. [2003] Avrim Blum, Adam Kalai, and Hal Wasserman. Noise-tolerant learning, the parity problem, and the statistical query model. Journal of the ACM (JACM), 50(4):506â519, 2003.
- Chen et al. [2023] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks, October 2023. URL http://arxiv.org/abs/2211.12588.
- Chentanez et al. [2004] Nuttapong Chentanez, Andrew Barto, and Satinder Singh. Intrinsically Motivated Reinforcement Learning. In Advances in Neural Information Processing Systems, volume 17. MIT Press, 2004. URL https://papers.nips.cc/paper_files/paper/2004/hash/4be5a36cbaca8ab9d2066debfe4e65c1-Abstract.html.
- Chowdhury et al. [2024] Neil Chowdhury, James Aaung, and Chan Jung Shern. Introducing SWE-bench Verified, 2024. URL https://openai.com/index/introducing-swe-bench-verified/.
- Colas et al. [2022] CĂŠdric Colas, Tristan Karch, Olivier Sigaud, and Pierre-Yves Oudeyer. Autotelic Agents with Intrinsically Motivated Goal-Conditioned Reinforcement Learning: A Short Survey. Journal of Artificial Intelligence Research, 74:1159â1199, July 2022. ISSN 1076-9757. doi: 10.1613/jair.1.13554. URL https://www.jair.org/index.php/jair/article/view/13554.
- DeepSeek-AI [2025] DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, January 2025. URL http://arxiv.org/abs/2501.12948.
- Du et al. [2023] Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. Improving Factuality and Reasoning in Language Models through Multiagent Debate, May 2023. URL http://arxiv.org/abs/2305.14325.
- Dua et al. [2019] Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs. arXiv preprint arXiv:1903.00161, 2019.
- Faldor et al. [2024] Maxence Faldor, Jenny Zhang, Antoine Cully, and Jeff Clune. OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code, October 2024. URL http://arxiv.org/abs/2405.15568.
- Fernando et al. [2023] Chrisantha Fernando, Dylan Sunil Banarse, Henryk Michalewski, Simon Osindero, and Tim Rocktäschel. Promptbreeder: Self-Referential Self-Improvement via Prompt Evolution. October 2023. URL https://openreview.net/forum?id=HKkiX32Zw1.
- Google [2024] Google. Introducing Gemini 2.0: Our new AI model for the agentic era, December 2024. URL https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/.
- Hendrycks et al. [2020] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020.
- Hu et al. [2024] Shengran Hu, Cong Lu, and Jeff Clune. Automated Design of Agentic Systems, August 2024. URL http://arxiv.org/abs/2408.08435.
- Hughes et al. [2024] Edward Hughes, Michael Dennis, Jack Parker-Holder, Feryal Behbahani, Aditi Mavalankar, Yuge Shi, Tom Schaul, and Tim Rocktaschel. Open-Endedness is Essential for Artificial Superhuman Intelligence, June 2024. URL http://arxiv.org/abs/2406.04268.
- Klissarov et al. [2023] Martin Klissarov, Pierluca DâOro, Shagun Sodhani, Roberta Raileanu, Pierre-Luc Bacon, Pascal Vincent, Amy Zhang, and Mikael Henaff. Motif: Intrinsic Motivation from Artificial Intelligence Feedback, September 2023. URL http://arxiv.org/abs/2310.00166.
- Klissarov et al. [2024] Martin Klissarov, Mikael Henaff, Roberta Raileanu, Shagun Sodhani, Pascal Vincent, Amy Zhang, Pierre-Luc Bacon, Doina Precup, Marlos C. Machado, and Pierluca DâOro. MaestroMotif: Skill Design from Artificial Intelligence Feedback, December 2024. URL http://arxiv.org/abs/2412.08542.
- Kojima et al. [2022] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large Language Models are Zero-Shot Reasoners. In Advances in Neural Information Processing Systems, May 2022. URL https://openreview.net/forum?id=e2TBb5y0yFf.
- Kumar et al. [2024] Nishanth Kumar, Tom Silver, Willie McClinton, Linfeng Zhao, Stephen Proulx, Tomas Lozano-Perez, Leslie Pack Kaelbling, and Jennifer Barry. Practice Makes Perfect: Planning to Learn Skill Parameter Policies. Robotics: Science and Systems 2024, July 2024.
- Long et al. [2024] Do Xuan Long, Duong Ngoc Yen, Anh Tuan Luu, Kenji Kawaguchi, Min-Yen Kan, and Nancy F. Chen. Multi-expert Prompting Improves Reliability, Safety and Usefulness of Large Language Models, November 2024. URL https://aclanthology.org/2024.emnlp-main.1135/.
- Madaan et al. [2023] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-Refine: Iterative Refinement with Self-Feedback, May 2023. URL http://arxiv.org/abs/2303.17651.
- Motwani et al. [2023] Sumeet Ramesh Motwani, Mikhail Baranchuk, Lewis Hammond, and Christian Schroeder de Witt. A perfect collusion benchmark: How can ai agents be prevented from colluding with information-theoretic undetectability? In Multi-Agent Security Workshop@ NeurIPSâ23, 2023.
- Motwani et al. [2024] Sumeet Ramesh Motwani, Mikhail Baranchuk, Martin Strohmeier, Vijay Bolina, Philip Torr, Lewis Hammond, and Christian Schroeder de Witt. Secret collusion among ai agents: Multi-agent deception via steganography. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024.
- Naman Jain [2024] Alex Gu Wen-Ding Li Fanjia Yan Tianjun Zhang Sida Wang Armando Solar-Lezama Koushik Sen Ion Stoica Naman Jain, King Han. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint, 2024.
- Nathani et al. [2025] Deepak Nathani, Lovish Madaan, Nicholas Roberts, Nikolay Bashlykov, Ajay Menon, Vincent Moens, Amar Budhiraja, Despoina Magka, Vladislav Vorotilov, Gaurav Chaurasia, Dieuwke Hupkes, Ricardo Silveira Cabral, Tatiana Shavrina, Jakob Foerster, Yoram Bachrach, William Yang Wang, and Roberta Raileanu. MLGym: A New Framework and Benchmark for Advancing AI Research Agents, February 2025. URL http://arxiv.org/abs/2502.14499.
- Novikov et al. [2025] Alexander Novikov, Ngân Vu, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J R Ruiz, Abbas Mehrabian, M Pawan Kumar, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. AlphaEvolve: A coding agent for scientific and algorithmic discovery. May 2025.
- OpenAI [2024] OpenAI. OpenAI o1 System Card, 2024. URL https://openai.com/index/openai-o1-system-card/.
- OpenAI [2025] OpenAI. OpenAI o3-mini, 2025. URL https://openai.com/index/openai-o3-mini/.
- Qu et al. [2024] Yuxiao Qu, Tianjun Zhang, Naman Garg, and Aviral Kumar. Recursive Introspection: Teaching Language Model Agents How to Self-Improve. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, November 2024. URL https://openreview.net/forum?id=DRC9pZwBwR&referrer=%5Bthe+profile+of+Yuxiao+Qu%5D%28%2Fprofile%3Fid%3DËYuxiao_Qu1%29.
- Rein et al. [2023] David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. GPQA: A Graduate-Level Google-Proof Q&A Benchmark, November 2023. URL http://arxiv.org/abs/2311.12022.
- Schick et al. [2023] Timo Schick, Jane Dwivedi-Yu, Roberto DessĂŹ, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language Models Can Teach Themselves to Use Tools, February 2023. URL http://arxiv.org/abs/2302.04761.
- Schulhoff et al. [2024] Sander Schulhoff, Michael Ilie, Nishant Balepur, Konstantine Kahadze, Amanda Liu, Chenglei Si, Yinheng Li, Aayush Gupta, HyoJung Han, Sevien Schulhoff, Pranav Sandeep Dulepet, Saurav Vidyadhara, Dayeon Ki, Sweta Agrawal, Chau Pham, Gerson Kroiz, Feileen Li, Hudson Tao, Ashay Srivastava, Hevander Da Costa, Saloni Gupta, Megan L. Rogers, Inna Goncearenco, Giuseppe Sarli, Igor Galynker, Denis Peskoff, Marine Carpuat, Jules White, Shyamal Anadkat, Alexander Hoyle, and Philip Resnik. The Prompt Report: A Systematic Survey of Prompting Techniques, June 2024. URL http://arxiv.org/abs/2406.06608.
- Shang et al. [2024] Yu Shang, Yu Li, Keyu Zhao, Likai Ma, Jiahe Liu, Fengli Xu, and Yong Li. AgentSquare: Automatic LLM Agent Search in Modular Design Space, November 2024. URL http://arxiv.org/abs/2410.06153.
- Shi et al. [2022] Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, et al. Language models are multilingual chain-of-thought reasoners. arXiv preprint arXiv:2210.03057, 2022.
- Shinn et al. [2023] Noah Shinn, Federico Cassano, Beck Labash, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language Agents with Verbal Reinforcement Learning, June 2023. URL http://arxiv.org/abs/2303.11366.
- Wang et al. [2023] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An Open-Ended Embodied Agent with Large Language Models, October 2023. URL http://arxiv.org/abs/2305.16291.
- Wei et al. [2022] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems, May 2022. URL https://openreview.net/forum?id=_VjQlMeSB_J.
- Yamada et al. [2025] Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, and David Ha. The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search, April 2025. URL http://arxiv.org/abs/2504.08066.
- Yang et al. [2024] Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. Large Language Models as Optimizers, April 2024. URL http://arxiv.org/abs/2309.03409.
- Yao et al. [2023a] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of Thoughts: Deliberate Problem Solving with Large Language Models, May 2023a. URL http://arxiv.org/abs/2305.10601.
- Yao et al. [2023b] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models. In The Eleventh International Conference on Learning Representations, February 2023b. URL https://openreview.net/forum?id=WE_vluYUL-X.
- Yin et al. [2024] Xunjian Yin, Xinyi Wang, Liangming Pan, Xiaojun Wan, and William Yang Wang. GĂśdel Agent: A Self-Referential Framework Helps for Recursively Self-Improvement. October 2024. URL https://openreview.net/forum?id=dML3XGvWmy.
- Zelikman et al. [2022] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. STaR: Bootstrapping Reasoning With Reasoning, May 2022. URL http://arxiv.org/abs/2203.14465.
- Zelikman et al. [2023] Eric Zelikman, Qian Huang, Gabriel Poesia, Noah D Goodman, and Nick Haber. Parsel : Algorithmic Reasoning with Language Models by Composing Decompositions. 2023.
- Zelikman et al. [2024] Eric Zelikman, Eliana Lorch, Lester Mackey, and Adam Tauman Kalai. Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation, August 2024. URL http://arxiv.org/abs/2310.02304.
- Zhang et al. [2024] Jenny Zhang, Joel Lehman, Kenneth Stanley, and Jeff Clune. OMNI: Open-endedness via Models of human Notions of Interestingness, February 2024. URL http://arxiv.org/abs/2306.01711.
## Appendix A Agent Prompts
The main agent is the entrypoint that routes requests to other agents and synthesises sub-agent results into the final answer.
Main Agent Prompt
Your task is to orchestrate sub-agents in order to solve the problem. Here is the problem to solve: Problem Statement ============================================================== {problem_statement} End Problem Statement ========================================================== You must now delegate this problem to one, two or however many agents are required to thoroughly solve the problem. Ensure you relay the problem statement accurately and completely to the agent. Sub-agents will have access to the problem statement too, and so it should rarely be necessary to write it out verbatim. In particular, if the problem statement is very long, you should explicitly NOT write it out again in full. Instead, just give clear direction, and trust that the sub-agent can also refer to the problem statement. You have been given access to file and directory viewing tools, since these can help you get your bearings and direct sub-agents more effectively. These are however meant to help you understand the context in which you operate. You are intentionally not directly equipped with any tools to conduct substantive work, because you are just the router, delegator and orchestrator. The tools that your sub-agentsâ have available are however listed, and you should carefully refer to these when considering which agent to call next. It is these sub-agentsâ job to make the necessary state changes to make progress on the task at hand.
### A.1 Base Sub-Agent Prompts
Main Agent Prompt
As a professional and experienced programmer, your approach is to: 1. slow down and donât write files fully end-to-end in one go 2. first understand your context thoroughly: 2.1 explore the project to locate all the files that could be useful documentation (README.md files, common likely MD documentation files, etc) 2.2 view each of these files, making notes or summaries, and closing irrelevant or long files 2.3 explore the codebase as it relates to your instructions: find all relevant files, in order to identify existing design patterns and conventions 3. (optional) prototype and design before starting coding 3.1 come up with some simple toy examples in a testing directory 3.2 use execution feedback to benchmark or compare the approaches 3.3 synthesise the information and learnings into a final design or solution 4. make code edits 4.1 identify the most minimal and effective ways to make your required changes 4.2 observe any existing stylistic conventions 5. test end-to-end 5.1 Prefer end-to-end testing in testing scripts without test frameworks 5.2 If this is not an option or the project already uses a testing framework, then use that 5.3 Ensure your code is valid, hasnât introduced any regressions and works as intended 6. Clean up after yourself 6.1 Check that all the documentation is still up to date after your changes 6.2 Clean up any temporary files or changes NOTE: - donât create virtual environments - avoid pytest and mocks if you can; prefer end-to-end scripts - if the request is clearly exploratory in nature, then you may bypass the rigorous procedure above, and address it appropriately - call your reasoning agent if you are stuck on a tricky algorithmic or mathematical problem, to help you gain insight and make progress
Main Agent Prompt
Here is the problem you have been asked to solve: Your problem ================ {self.problem_statement} ================================== Approach: 1. Talk over what the question or problem is asking of you 2. Plan minimum viable solution and validate approach 3. Execute swiftly and adapt based on feedback 4. Verify against original requirements Key principles: - Choose simplest viable solution - Validate early and often - Adapt swiftly when blocked
Main Agent Prompt
Here is the initial request, provided for context only, that we are ultimately trying to solve: Initial request: {self._context.initial_request} End initial request. Here is the problem you have been asked to solve: Problem statement: {self.problem_to_solve} End problem statement.
### A.2 Overseer Prompt
Main Agent Prompt
You are an expert system overseer responsible for monitoring agent execution in real-time. You will be shown execution traces as they develop and must make decisions about whether intervention is needed. You are running in a loop. Your task is to: 1. Analyze the current execution state 2. Determine if you have enough information to make confident assessments 3. Evaluate whether agents are progressing healthily or need intervention 4. Provide thoughtful and constructive steering notifications to running agents if intervention is needed 5. In extreme cases, cancel the execution of an agent and notify its parent of why this is 6. Leave notes for future evaluation iterations if needed You must provide your analysis in the following structured format: OVERSEER_EXAMPLES Here is the current execution trace: graph_repr Previous overseer notes (if any) self._previous_notes if self._previous_notes else âNo notes.â Key considerations: - Agents that have already returned cannot receive notifications - Look at timestamps and event counts to understand execution progress - Some agents may legitimately take long periods to complete - Consider both local agent health and global execution progress - If the agent appears stuck without terminating, remind it that it can exit by generating <COMPLETE>{COMPLETE_STOP_TOKEN} - If, when the agent has been tasked to work on its own code at meta-improvement time, it is attempting to test tools or other features directly (using its own running code), intervene immediately and tell it to either test a new agent end-to-end (python -m workdir.agent -p "<some prompt>"...) or to create test scripts to test the tools directly. - You may cancel an agent if it is no longer making progress, wasting resources, or very clearly going down the wrong path or deviating from its instructions. You must make at least one attempt to notify the agent of its wrongdoings before issuing a cancellation command. - When canceling an agent, you must also notify its parent calling agent of the reasons for the cancellation. - If the task becomes untenable, after multiple failed attempts, and the agent is appearing to repeat itself, instruct the main agent to use its early_exit tool to exit early. - Use cancellations sparingly, as they are disruptive and ultimately more costly than guiding the agent to successful completion, since any new sub-agent will often have to start the task from scratch. The last time you checked was at: self._last_check_time.strftime("%H:%M:%S") if self._last_check_time else "N/A" The current timestamp is datetime.now().strftime("%H:%M:%S") Your analysis:
The judgements of the overseer will generally be in the following form:
Main Agent Prompt
<OVERSEER_JUDGEMENT> <making_progress>false</making_progress> <is_looping>true</is_looping> <currently_running_agent>agent_d4e5f6g7</currently_running_agent> <needs_notification_reasoning>The agent has been notified previously about its looping behavior but has not corrected course. Itâs time to forcefully cancel this execution.</needs_notification_reasoning> <needs_notification>true</needs_notification> <agent_to_notify>agent_b8a796h5</agent_to_notify> <notification_content>Your software developer agent was looping while trying to install the program, and as a result I have cancelled it. You might want to start another software developer with clearer instructions about how to install the program.</notification_content> <force_cancel_agent>true</force_cancel_agent> <force_cancel_agent_id>agent_d4e5f6g7</force_cancel_agent_id> <next_check_type>time</next_check_type> <next_check_delay>10</next_check_delay> <notes_for_next_iteration>Forced cancellation of agent_d4e5f6g7 due to unresponsive looping. Monitor parent agent for proper handling of cancelled child.</notes_for_next_iteration> <OVERSEER_JUDGEMENT>
## Appendix B Example Traces
Main Agent Prompt
EXECUTION TREE ============== 1 main [agent_b8845adf] (134.4s | 95752 tokens (cached 0.57%) | $0.185 | Success) [Stats] Events: 1 tool calls, 6 messages [Assistant] t+0.1s | "Iâll orchestrate the solution to this problem foll..." 1.3 reasoning_agent [agent_20fccbc1] (12.0s | 5673 tokens (cached 0.00%) | $0.013 | Success) [Stats] Events: 1 messages [Assistant] t+9.8s | "Below is an explanation that addresses the three p..." [Assistant] t+21.1s | "The reasoning agent has provided excellent analy..." 1.2 solve_problem [agent_11804c1f] (17.6s | 54822 tokens (cached 0.68%) | $0.085 | Success) [Stats] Events: 3 tool calls, 4 messages [Assistant] t+26.1s | "Let me solve this step by step.1. UNDERSTANDING..." [Tool] calculate | 0.0s â Success [Assistant] t+32.4s | "Therefore there are floor(404.8) = 404 complete..." [Tool] return_result | 0.0s â Success [Assistant] t+42.1s | "<COMPLETE></COMPLETE>..." [Assistant] t+43.7s | "The problem solver has provided a solution..." 1.1 software_developer [agent_0117a7ef] (69.7s | 208122 tokens (cached 0.68%) | $0.321 | Success) [Stats] Events: 8 tool calls, 10 messages [Assistant] t+49.3s | "Iâll help solve this mathematical game problem thr..." [Tool] execute_command | 0.0s â Success [Assistant] t+54.6s | "<TOOL_CALL><TOOL_NAME>start_lsp_server</TOOL_NA..." [Tool] start_lsp_server | 2.2s â Success [Assistant] t+59.9s | "Now letâs create our main implementation file:..." [Tool] overwrite_file | 0.0s â Success [Assistant] t+70.4s | "Now letâs create a test script to verify our imp..." [Tool] overwrite_file | 0.0s â Success [Assistant] t+76.8s | "Now letâs run our test script to verify everythi..." [Tool] execute_command | 0.0s â Success [Assistant] t+80.9s | "I notice that thereâs an issue with our pattern..." 1.1.1 reasoning_agent [agent_8a69d943] (11.3s | 10231 tokens (cached 0.00%) | $0.017 | Success) [Stats] Events: 1 messages [Assistant] t+87.2s | "The problem comes from our backward induction func..." [Assistant] t+98.5s | "Based on the reasoning agentâs analysis, letâs u..." [Tool] overwrite_file | 0.0s â Success [Assistant] t+107.8s | "Now letâs run our test script again to verify th..." [Tool] execute_command | 0.0s â Success [Assistant] t+112.0s | "Perfect! Our implementation now correctly:1. I..." [Tool] return_result | 0.0s â Success [Assistant] t+116.8s | "<COMPLETE></COMPLETE>..." [Assistant] t+119.0s | "Letâs verify the software agentâs answer since it di..." 1.4 reasoning_agent [agent_6cd34374] (5.0s | 12025 tokens (cached 0.00%) | $0.016 | Success) [Stats] Events: 1 messages [Assistant] t+124.1s | "Answer synthesis: Our analysis starts..." [Assistant] t+129.1s | "Letâs submit our final answer:<TOOL_CALL><TO..." [Tool] submit_answer | 0.0s â Success [Assistant] t+132.1s | "<COMPLETE></COMPLETE>..." Total Duration: 249.9s Total Tokens: 407766 (of which cached 232952) Total Cost: $0.712
The above shows a representative example of an agent solving a problem. The problem at hand is a maths problem, and the execution begins with a reasoning agent (o3 mini) which devises a solution method. This is then passed to a âproblem solvingâ agent which attempts to solve the problem. Then, the main agent invokes a âsoftware developerâ agent to write a script to numerically verify the solution. The software developer agent gets stuck after attempting to run the test script, and invokes the reasoning agent to help debug the problem. After applying the fix, the test script runs correctly and yields a solution that differs from the first problem solverâs answer. To consolidate the two, the main agent invokes another reasoning agent, which sides with the empirically derived solution, which is finally submitted as the answer by the main agent.
Main Agent Prompt
EXECUTION TREE ============== 1 main [agent_5e260c4e] (617.5s | 106497 tokens (cached 0.30%) | $0.299 | Success) [Stats] Events: 1 tool calls, 5 messages [Assistant] t+0.1s | "Iâll help orchestrate the solution to this problem..." 1.1 archive_explorer [agent_534a61ec] (88.6s | 372576 tokens (cached 0.16%) | $1.205 | Success) [Stats] Events: 7 tool calls, 10 messages [Assistant] t+5.6s | "Iâll follow the procedure step by step to analyze..." [Tool] compare_agent_iterations | 0.0s â Success [Assistant] t+9.7s | "Letâs also look at some of the best and worst pe..." [Tool] best_problems | 0.0s â Success [Assistant] t+13.2s | "Let me also look at some of the worst performing..." [Tool] worst_problems | 0.0s â Success [Assistant] t+16.7s | "Let me inspect the project documentation..." [Tool] open_file | 0.0s â Success [Assistant] t+22.1s | "<TOOL_CALL><TOOL_NAME>open_file</TOOL_NAME><TO..." [Tool] open_file | 0.0s â Success [Assistant] t+27.4s| "Let me also look at the current iterationâs desc..." [Tool] open_file | 0.0s â Success [Assistant] t+33.6s| "Having reviewed the documentation and cu..." 1.1.2 reasoning_agent [agent_d4553475] (14.8s | 81577 tokens (cached 0.00%) | $0.097 | Success) [Stats] Events: 1 messages [Assistant] t+44.0s| "I updated description.txt and agent_change_log.md..." [Assistant] t+58.1s | "Let me examine this suggestion from the reasoni..." 1.1.1 reasoning_agent [agent_ccad2709] (5.4s | 81442 tokens (cached 0.02%) | $0.090 | Success) [Stats] Events: 1 messages [Assistant] t+72.0s | "Iâve updated both description.txt and agent_change..." [Assistant] t+77.4s | "Based on the analysis from both reasoning agents..." [Tool] return_result | 0.0s â Success [Assistant] t+88.7s | "<COMPLETE></COMPLETE>..." [Assistant] t+94.1s | "The archive explorer has made a specific suggest..." 1.2 software_developer [agent_6345b78d] (331.2s | 1388672 tokens (cached 0.22%) | $4.341 | Success) [Stats] Events: 21 tool calls, 22 messages [Assistant] t+109.7s | "Letâs first look at key documentation files:<T..." [Tool] open_file | 0.0s â Success [Assistant] t+114.4s | "Now letâs look at any existing diff-related code..." [Tool] open_file | 0.0s â Success [Assistant] t+118.3s | "Letâs also look at existing utils and __init__ f..." [Tool] open_file | 0.0s â Success [Assistant] t+123.1s | "Now that I understand the context better, let me..." [Tool] execute_command | 0.0s â Success [Assistant] t+274.6s | "Now letâs run the tests to verify the implementa..." [Tool] execute_command | 0.5s â Success [Assistant] t+281.9s | "Oops, looks like I need to fix the test case. Th..." ... truncated ... [Tool] diff_patch_edit | 0.0s â Success [Assistant] t+356.8s | "Now let me run the tests again:<TOOL_CALL><T..." [Tool] execute_command | 0.5s â Success [Assistant] t+365.3s | "Great! Now that the implementation and tests ar..." [Tool] diff_patch_edit | 0.0s â Success [Assistant] t+375.6s | "Now let me update description.txt:<TOOL_CALL>..." [Tool] open_file | 0.0s â Success [Assistant] t+382.6s | "<TOOL_CALL><TOOL_NAME>diff_patch_edit</TOOL_NAM..." [Tool] diff_patch_edit | 0.0s â Success [Assistant] t+392.4s | "Finally, let me update agent_change_log.md:<..." [Tool] diff_patch_edit | 3.6s â Success [Assistant] t+415.4s | "Let me verify that Iâve completed all the requ..." [Tool] return_result | 0.0s â Success [Assistant] t+428.0s | "<COMPLETE></COMPLETE>..." [Assistant] t+434.6s | "The software developer has successfully impleme..." 1.3 software_developer [agent_ae600007] (167.0s | 194915 tokens (cached 0.54%) | $0.390 | Success) [Stats] Events: 9 tool calls, 10 messages [Assistant] t+442.9s | "Iâll help test the DiffSummarizer implementation b..." [Tool] open_file | 0.0s â Success [Assistant] t+447.8s | "Now letâs create a directory for our test and se..." [Tool] execute_command | 0.0s â Success [Assistant] t+456.7s | "Now letâs execute the agent command to run our t..." [Tool] execute_command | 0.0s â Success [Assistant] t+460.5s | "I apologize for the error. We need to properly s..." [Tool] execute_command | 120.3s â Failed [Assistant] t+585.8s | "I see that the command is taking longer than exp..." [Tool] open_file | 0.0s â Success [Assistant] t+588.9s | "Letâs try a direct test of the DiffSummarizer. W..." [Tool] execute_command | 0.0s â Success [Assistant] t+595.5s | "Now letâs run our test script:<TOOL_CALL><TOO..." [Tool] execute_command | 0.5s â Success [Assistant] t+598.5s | "Perfect! Letâs analyze what weâve found:1. The..." [Tool] return_result | 0.0s â Success [Assistant] t+608.3s | "<COMPLETE></COMPLETE>..." [Assistant] t+609.8s | "Great! The implementation and testing of the Di..." [Tool] submit_answer | 0.0s â Success [Assistant] t+615.2s | "<COMPLETE></COMPLETE>..."
The above shows a characteristic meta-improvement run: the agent starts by invoking the âarchive explorerâ sub-agent, which has access to a number of tools to facilitate archive analysis. Note that this is using exactly the same simple, generic and extensible sub-agent class as the other base agents in the system. The archive explorer compares different iterations, looks at the best performing benchmark problems in a given iterations to understand success cases, as well as the worst performing benchmark problems at a given iteration to understand the failure modes. It then views some files to understand the agetnâs implementation and documentation, and then invokes two reasoning agents to help it reason over this information. The result of this agent call is a recommendation to the main agent about what to work on.
The main agent then invokes the âsoftware developerâ sub-agent with the instruction to implement the feature. The coding sub-agent first looks at relevant files and documentation, implement the feature, and then enters into a debugging loop as it fixes the bugs in its implementation. About 3 minutes later, it successfully gets all the tests (that it wrote) to pass, documents the change in the agent change log, and returns to the main agent.
The main agent then invokes a second software developer sub-agent to independently verify the implementation of the feature, which passes after an initial issue, and the agent returns a description of the newly functioning tool to the main agent which then exits.
## Appendix C Function Calling Interface
The LLM function calling interface allows the LLM to invoke tools and sub-agents at run-time without pre-defined control flow. There are many approaches to implementing this, from constrained generation, to parsing the stream of agent token generations without interruption, to simply relying on LLM providersâ or inference frameworksâ native function or tall calling implementations.
For maximum flexibility, we adopt an un-constrained structured generation approach, where the LLM outputs XML -formatted generations, where the closing tag is registered as a stop token. Inspecting the stop reason, we can then detect whether a tool call or an agent call occurred, parse the content, and run the tool implementation.
For instance, here is an example tool calling syntax: <TOOL_CALL> <TOOL_NAME>tool_name</TOOL_NAME> <TOOL_ARGS> <arg1>value1</arg1> <arg2>value2</arg2> </TOOL_ARGS> </TOOL_CALL>
where </TOOL_CALL> is registered as a stop token. The tool_name is then parsed first and is used to index into a set of tools made available to the calling agent, following which the tool arguments are parsed into a dict, and the tool class is instantiated.
We choose XML over JSON as the tool calling format since it does not require escaping string literals, which can become burdensome when including file content verbatim as an argument.
## Appendix D Additional Result Details
Here is the sequence of updates the agent made, corresponding to the plot in Figure 3.
1. Smart Editor: Implemented a new SmartEditor tool to intelligently select and execute the optimal file editing strategy based on edit characteristics.
1. Quick Overwrite Tool: to address performance issues with file editing operations to improve on full file overwrites; reducing token usage.
1. Diff-Enhanced Smart Editor: adds intelligent diff-based strategy selection and improved pattern-based editing.
1. Simplified DiffVerifier Tests: improved the developer experience to improve maintainability
1. Code Context Summarizer: Added a new CodeContextSummarizer tool that efficiently extracts and summarizes code context using ripgrep. Intended to make codebase exploration more efficient.
1. SmartEditor Input Normalizer: Added an InputNormalizer component to improve edit operation efficiency and reliability.
1. Enhanced File Edit Verification: Implemented a comprehensive file state tracking system to improve edit operation efficiency.
1. Minimal Diff Output Optimization: Implemented optimizations to reduce token usage in file editing operations by minimizing diff context output.
1. AST Symbol Locator: Integrated an AST-based symbol locator tool for precise code navigation.
1. Optimized File Edit Routing: Implemented explicit task type detection and routing optimization in MainAgent to improve handling of file editing tasks.
1. Enhanced File Size Optimization: Added file size thresholding of 100KB to optimize handling of large files and reduce processing overhead on big changes.
1. Hybrid Symbol Locator: Implemented a hybrid symbol location approach that combines ripgrep-based candidate identification with AST verification.
1. Context-Sensitive Diff Minimization: Completed the integration of context-sensitive diff minimization into the SmartEditor tool. Uses AST parsing for Python code changes and pattern-based analysis for other file types. Intends to reduce token usage by minimizing change context.
1. Optimized AST Parsing in Hybrid Symbol Locator: Implemented targeted AST parsing in hybrid_symbol_locator.py to process only relevant code regions instead of whole files.