Evolutionary optimization where the genome is just a directory of files

What it does

AntOmniEvo is an auto-evolution framework for AI agent systems and other tunable software. You define five things — how to run your system, how to score it, eval data, a schema mapping your tunable parts onto a directory of files, and an initial version of those files. The framework then runs a concurrent evolution loop where a coding-agent Proposer reads failure trajectories and rewrites the artifacts, the way a human engineer would edit code. What you get back is the best candidate’s tunable artifacts: a real directory of files you can diff, review, and deploy.

The interesting bit

The abstraction is the trick: anything expressible as a directory of files with a repeatable, reasonably-cheap evaluation becomes optimizable — agent skills, workflow configs like a pipeline.json plus node scripts, or a single-file algorithm. The system under optimization doesn’t even need to contain an LLM; only the proposer does. Every edit is attributed to the failure evidence that motivated it, so the change lineage reads like a code review, not a mystery mutation.

Key highlights

• Strict division of labor: you define the objective, the framework handles scheduling, budgets, selection/elimination, and persistence.
• All candidates, runs, analyses, and changelogs persist to a CandidateStore — interruptible and resumable.
• Works on agent skill/harness/memory directories, workflow pipelines, and single-file algorithms alike.
• Backed by a paper (“Mara Chain”) on treating failure as a stepping stone for auto-evolution.
• Bilingual docs (English and Chinese) covering quickstart, extensibility, checkpoint resume, and a visualizer.

Caveats

• The README doesn’t quantify results or benchmarks — you’ll need the paper or your own runs to judge effectiveness.
• “Optimizes anything” really means “anything that fits the directory-of-files plus repeatable-eval mold”; systems with expensive or non-deterministic evaluation may strain the loop.

Verdict

Worth a look if you maintain an agent or pipeline whose prompts, skills, or configs drift out of tune and you have a cheap, repeatable eval. If your system can’t be scored automatically, this isn’t your tool — the whole loop runs on that evaluator.