A Self-Improving Coding Agent

2025-04-21

AI Summary

This work addresses the challenge of enabling coding agents to achieve closed-loop autonomous evolution using only foundational capabilities—code execution, tool invocation, self-reflection, and iterative code rewriting—without external human intervention or predefined update rules. We propose the first large language model–based programming agent capable of fully automated, open-ended self-modification, wherein the agent continuously refactors its own implementation to improve task performance. Our core contribution is formalizing self-modification as an executable, verifiable software engineering process—shifting embodied agents from reactive to evolutionary paradigms. Empirically, the agent achieves 17–53% absolute performance gains on the SWE-Bench Verified subset, and demonstrates consistent improvements on LiveCodeBench and synthetic agent benchmarks, validating both the efficacy and generalizability of self-evolution.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Cognitive Modeling & Cognitive Systems: Agent Architectures
• Multiagent Systems: Agent-Based Simulation and Emergent Behavior

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

We demonstrate that an LLM coding agent, 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 provides a reference agent framework for those seeking to post-train LLMs on tool use and other agentic tasks.

Problem

• Enabling LLM coding agents to autonomously edit themselves
• Improving performance on benchmark coding tasks
• Advancing automated design of agentic systems

Innovation

• LLM coding agent autonomously edits itself
• Performance gains up to 53% on benchmarks
• Reference framework for post-training agentic tasks