AntOmniEvo

AntOmniEvo, from Ant Research, is an auto-evolution framework with a strict division of labor. You define five things once: how to run your system on one instance, an evaluator that scores output from 0 to 1, train/validation data, a schema mapping the tunable files, and the initial artifacts. The framework runs the evolution loop, and a coding-agent Proposer does the actual editing.

Tunable artifacts are real directories: an agent's SKILL.md with references and scripts, a workflow's pipeline.json with node code, or a single-file algorithm. The Proposer reads failure trajectories, picks which file to change and writes a new candidate, while the framework handles scheduling, budgets, selection and persistence in a resumable CandidateStore.

The output is the best candidate's files, diffable and deployable, with a change lineage linking each edit to the failure that motivated it. It needs a repeatable, reasonably cheap evaluation to work. The bundled skill example is in Chinese. It accompanies the paper Mara Chain and is licensed Apache-2.0.