ant-research/AntOmniEvo

What it solves

AntOmniEvo is an auto-evolution framework designed to automate the iterative process of optimizing AI agent systems, workflows, and algorithms. It replaces the manual trial-and-error process of editing prompts, configuration files, and code to improve performance based on evaluation results.

How it works

The framework treats the tunable parts of a system—such as prompts, skill files, and pipeline configurations—as "tunable artifacts" (a directory of files). It then runs a concurrent evolution loop consisting of three roles:

1. The User: Defines the system execution logic, the scoring evaluator, the evaluation data, and the initial artifacts.
2. The Framework: Manages the control loop, including scheduling, candidate selection, elimination, and persistence of all runs and changelogs.
3. The AI Proposer: A coding agent that analyzes failure trajectories from the evaluation and rewrites the tunable artifacts to fix errors, mimicking how a human engineer would edit code.

Who it’s for

It is intended for developers building AI agents, retrieval-augmented generation (RAG) pipelines, or single-file algorithms who have a repeatable and reasonably cheap way to evaluate their system's performance.

Highlights

• Artifact-based optimization: Treats any system representable as a directory of files as a genome for evolution.
• Failure-driven evolution: Uses AI agents to analyze specific failure cases to motivate changes.
• Transparent lineage: Provides a change lineage that attributes every edit to the failure evidence that triggered it.
• Resilient execution: Includes a CandidateStore for interruptible and resumable checkpoints and crash recovery.