Prompt and agent optimizers usually mutate something narrow — a constrained parameter set, a DSL, a prompt string — while the thing an agent system actually exposes is a directory of files. AntOmniEvo from Ant Research takes that literally: the optimisation target is the tunable artifact, "a real directory of files", and mutation is a coding agent adding, deleting or editing files inside it. Optimisation becomes plain file editing. The shapes it names are an agent's SKILL.md plus references and scripts, a workflow's pipeline.json plus nodes/*.py, and a single-file algorithm plus its description. The system under optimisation need not contain an LLM at all — the System interface only sees "tunable artifacts + data → trajectory + output" — but the Proposer must be one, because editing files requires coding capability.

The division of labour is the design. You define five things once: a System (how to run on one eval instance), an Evaluator (score 0–1, whose scoring criteria is the objective), train/val eval data, a TunableArtifactSchema mapping your tunables onto a file tree with a purpose per file, and the initial artifacts. The framework controls the loop — scheduling, budgets, selection and elimination, persistence — with every candidate, run, analysis and changelog written to a CandidateStore so a run is interruptible and resumable. The AI works: a coding-agent Proposer reads failure trajectories, locates the file to edit, and lands a structured ArtifactAction{file, operation, artifact_issue, change, resolves[]} that must cite the failed-trajectory observation it fixes. It runs in two phases so raw runs never reach the mutator: phase one distils a failed run into a structured RunAnalysis, phase two deduplicates and resolves contradictions across instances, behind a self-check gate and an enforced append-only changelog.jsonl that a child inherits from its parent. What you get back is the best candidate's directory — diffable, reviewable, deployable — with a lineage attributing every edit to the evidence that motivated it.

Three mechanisms carry the loop. Multi-slot concurrent evolution lets up to num_proposals candidates evolve simultaneously over a shared pool and state machine, where the repo notes prior work (GEPA, ACE, AHE) is serial. Mara-chain deep repair handles a batch that misses its threshold without stopping at one natural-language reflection: the failed chain's per-instance score history, attributed changelog and prior analyses are pre-loaded into the prompt, forcing a five-method diagnosis — causal tracing, success extraction, oscillation detection, inconsistency discovery, evidence corroboration — for up to max_reflection_iterations rounds, accepting the best chain member that beats the original parent. Micro-pareto selection prunes by per-instance Pareto weak dominance with a top-N truncation fallback. Supporting machinery: four orthogonal hard budget caps (iterations, rollouts, tokens, wall clock) with soft-stop-and-drain semantics, checkpoint resume via a candidate state machine, and strict validation-set isolation — val_system_run/ is off-limits to the proposer, hard-blocked by a runtime extension on the Pi path, so a training batch decides whether a mutation is worth keeping and the val set decides where the candidate sits in the population.

Two things to hold it by. The documentation concedes its own novelty boundary unprompted: "Pareto selection itself is adopted from GEPA, not an AntOmniEvo contribution; AntOmniEvo's algorithmic deltas are concurrency and mara chains." And the README reports no benchmark numbers at all for the framework — the evidence is the design, not a scoreboard; the companion paper Mara Chain: Rethinking Failure as a Stepping Stone for AI System Auto-Evolution (arXiv 2609.35855, 25 September 2026, Yubin Lyu et al.) arrived ten days after the repository. The stated non-fits are useful: systems with no tunable artifact (a few scalar parameters belong in black-box tuning), non-repeatable or extremely expensive evaluation, and strict-compliance output, since the proposer edits files in a sandbox whose permissions you must audit yourself. Python with a React and Flask run visualizer, Apache-2.0, ~1,140 GitHub stars and 101 forks at filing — adjacent to Microsoft's AutoSaddler, which attacks the same trace-to-patch loop but scoped to a coding-agent harness.

Library

Language Python
License Apache-2.0
agentsagent-harnessagenticframeworkself-improvementopen-source

Related