1.2K ·GitHub
What is this?
AntOmniEvo is an auto-evolution framework: you give it your system's tunable artifacts (skills, prompts, workflow configs, or code) plus a way to score them, and it runs a concurrent evolution loop where a coding-agent 'Proposer' rewrites those files to produce better-performing candidates.
AntOmniEvo is an auto-evolution framework: you give it your system's tunable artifacts (skills, prompts, workflow configs, or code) plus a way to score them, and it runs a concurrent evolution loop where a coding-agent 'Proposer' rewrites those files to produce better-performing candidates.Is this a framework I plug into my own system, or a standalone app?It is a framework, not a standalone app or a plain library you call. You provide a System runner, an Evaluator, eval data, and a TunableArtifactSchema, and the framework drives the optimization loop and the coding-agent Proposer for you.
It is a framework, not a standalone app or a plain library you call. You provide a System runner, an Evaluator, eval data, and a TunableArtifactSchema, and the framework drives the optimization loop and the coding-agent Proposer for you.How do I get it running on my own system?Follow docs/quickstart.md for install and quick-start steps: you wire up a System (how to run your system on one eval instance), an Evaluator (how to score output 0-1), eval data, and a TunableArtifactSchema describing the tunable file tree, then start the evolution loop.
Follow docs/quickstart.md for install and quick-start steps: you wire up a System (how to run your system on one eval instance), an Evaluator (how to score output 0-1), eval data, and a TunableArtifactSchema describing the tunable file tree, then start the evolution loop.What do I need to supply before I can start an evolution run?You need your own System, Evaluator, and eval data, plus a coding-agent that will act as the Proposer editing the artifacts — the framework itself doesn't ship an LLM or evaluator.
You need your own System, Evaluator, and eval data, plus a coding-agent that will act as the Proposer editing the artifacts — the framework itself doesn't ship an LLM or evaluator.Any catch I should know before relying on this?The system you're optimizing doesn't need to contain an LLM, but the Proposer that edits your files must be an agent, and the project's accompanying paper hasn't been published yet.
The system you're optimizing doesn't need to contain an LLM, but the Proposer that edits your files must be an agent, and the project's accompanying paper hasn't been published yet.Who is this meant for?It's aimed at teams who have an agent system, workflow, or algorithm that is already expressible as files and can be scored repeatably and cheaply — effectively giving them a round-the-clock team of algorithm engineers automating that tuning.
It's aimed at teams who have an agent system, workflow, or algorithm that is already expressible as files and can be scored repeatably and cheaply — effectively giving them a round-the-clock team of algorithm engineers automating that tuning.What exactly counts as a 'tunable artifact'?Anything expressible as a directory of files with a repeatable evaluation: an agent's SKILL.md plus references and scripts, a workflow's pipeline.json plus node code, or a single algorithm file with its description.
Anything expressible as a directory of files with a repeatable evaluation: an agent's SKILL.md plus references and scripts, a workflow's pipeline.json plus node code, or a single algorithm file with its description.How does the Proposer decide what to change?It reads failure trajectories from evaluation runs, locates which file is responsible for the failure, and lands a structured edit as a new candidate, the way a human engineer would edit code.
It reads failure trajectories from evaluation runs, locates which file is responsible for the failure, and lands a structured edit as a new candidate, the way a human engineer would edit code.Can I stop and resume a long evolution run?Yes — every candidate, run, analysis, and changelog is persisted to a CandidateStore, so runs are interruptible and resumable, including crash recovery.
Yes — every candidate, run, analysis, and changelog is persisted to a CandidateStore, so runs are interruptible and resumable, including crash recovery.What do I actually get back at the end?The best candidate's tunable artifacts as a real directory of files you can diff, review, and deploy, with a change lineage tracing every edit back to the failure evidence that caused it.
The best candidate's tunable artifacts as a real directory of files you can diff, review, and deploy, with a change lineage tracing every edit back to the failure evidence that caused it.What are similar projects?PocketFlow, Raven, rrsi, harness-sdk, spring-ai-alibaba, evolver, qwen-code, foreman
PocketFlow, Raven, rrsi, harness-sdk, spring-ai-alibaba, evolver, qwen-code, foremanWhich topics is it in?self-improvement, loop-engineering, llm-agent, self-evolving, agentic, optimization, agent-framework, workflow-automation, agent-orchestration, multi-agent, agents, coding-agent
self-improvement, loop-engineering, llm-agent, self-evolving, agentic, optimization, agent-framework, workflow-automation, agent-orchestration, multi-agent, agents, coding-agent