EvoMap/evolver — GEP-Powered Self-Evolution Engine for AI Agents
Self-evolving engine that encodes agent experience as Genes/Capsules under GEP — every behavioral change is gated, traced, and reusable like git commits.

⭐ 2,200
Self-evolving engine that encodes agent experience as Genes/Capsules under GEP — every behavioral change is gated, traced, and reusable like git commits.
The reason this repo is trending isn't only the star count. What matters is the gap it fills in the AI agent ecosystem — and right now that gap is contested every week by new entrants.
Background
Maintainer history, employer affiliation, and contributor mix are the first credibility signals. README opening with a working demo video usually means PoC works; long text-only README typically signals pre-demo stage.
Core Capability
The fundamental problem here is how LLM agents efficiently manage tokens, memory, and tool calls in long-horizon workloads. LangChain/LlamaIndex are strong on single-shot RAG/chains; they accumulate inefficiency in multi-step autonomous execution.
This repo combines (1) context compression, (2) self-eval, (3) tool-call abstraction. Surface value: cut token cost in half on equivalent tasks. Deeper value: reproducible agent execution logs.
Stack
- Language: Python
- License: GPL-3.0-or-later
Comparison
| Project | Stars | Daily | Differentiator |
|---|---|---|---|
| This repo | 2,200 | 161 | (per summary) |
| AutoGPT | 170k | 50 | full autonomy, low maturity |
| LangGraph | 10k+ | 60 | graph workflow |
| CrewAI | 28k | 100 | multi-agent |
Daily-stars velocity matters more than cumulative right now; the agent category is not winner-take-all — it segments by use-case.
Why Now
Agent stack splitting in two: official IDE integrations (Codex CLI, Claude Code) vs open-source core libraries. Enterprise PoC demand is pulling the latter back into focus to cut SaaS lock-in cost.
Quickstart
git clone https://github.com/EvoMap/evolver
cd $(basename https://github.com/EvoMap/evolver .git)
pip install -r requirements.txt
export OPENAI_API_KEY=...
python examples/quickstart.py
Common first pitfalls: Pydantic v2 + Python <3.11 mismatch; rate-limit hitting mid-demo (cap with --max_iterations 5).
Limits + Roadmap
Two clear limits today: (1) non-English workload validation thin; (2) enterprise SSO/audit logs missing. Latter is on the June roadmap; former is open-issue only.
Tomorrow Morning
- Devs:
git clone https://github.com/EvoMap/evolver, run quickstart, port one workload to compare token cost. - Founders/PM: ROI sim if migrating from OpenAI Assistants API to OSS backbone.
- Investors/General: Watch daily-stars next 7 days. >200/day = hype peak.
Sources
관련 기사

This AI Rewrites Its Own Code — MiniMax M2.7's Self-Evolution Experiment
MiniMax M2.7 autonomously improved itself over 100+ iterations, scoring 56.22% on SWE-Pro — near Claude Opus 4.6 levels — at 1/50th the price.

lsdefine/GenericAgent — Self-Evolving Agent with 6× Less Token Use
Self-evolving agent that grows its skill tree from a 3.3K-line seed and achieves full system control with 6× fewer tokens than baselines via a ~100-line core loop.

OpenAI Put a Terminal in Its API – From Model Company to Agent Platform
OpenAI's Responses API now includes Shell tool, hosted containers, Skills, and Context Compaction. An agent infrastructure that maintains accuracy across 5-million-token sessions.
AI 트렌드를 앞서가세요
매일 아침, 엄선된 AI 뉴스를 받아보세요. 스팸 없음. 언제든 구독 취소.
