AntOmniEvo

AntOmniEvo is an open-source auto-evolution framework from the ant-research organization on GitHub. Users define a system's tunable artifacts and an evaluator, and a coding-agent proposer edits those artifacts over repeated rounds. It delivers the best candidate as a directory of files that can be diffed, reviewed, and deployed.

What It Is

AntOmniEvo treats a system's tunable parts (skills, prompts, workflow configs, pipeline code) as the genome of an evolution loop. Anything that can be represented as a directory of files and evaluated repeatably can be optimized, so optimization becomes plain file editing. The system being optimized does not need to contain an LLM, but the proposer must be agents.

Documented target types include AI agent skill, harness, memory, and extension directories, workflows or pipelines (for example a retrieval DAG's pipeline.json plus node scripts), and single-file algorithms with their descriptions.

How the Division of Labor Works

The user supplies five things once: a System that runs one eval instance, an Evaluator that scores output from 0 to 1, train/validation eval data, a TunableArtifactSchema mapping the artifact directory, and the initial tunable artifacts.

The framework controls the loop: scheduling, budgets, selection and elimination, and persistence. Every candidate, run, analysis, and changelog is saved to a CandidateStore, so runs can be interrupted and resumed.

The coding-agent Proposer reads failure trajectories, locates the file to edit, and writes a structured change as a new candidate.

Outputs and Documentation

The result is the best candidate's artifacts, with a change lineage linking each edit to the failure evidence that motivated it. The repository includes docs on quick start, features, extensibility, system design, checkpoint resume, and a visualizer. It is released under the Apache License 2.0 and cites the paper Mara Chain on failure-driven AI system auto-evolution.