EvoMap/evolver — +1,131 ⭐ in a day, GEP makes agents actually evolve
EvoMap's evolver introduces the Genome Evolution Protocol (GEP), encoding agent experience as Genes and Capsules and evolving them through mutation and selection. 4,590 controlled trials show gene representation is the strongest path. Yesterday alone: +1,131 stars.
+1,131 ⭐ in a day
A 1,000+ daily-star jump on an agent repo is rare. evolver did it. v1.28.0, the EvoMap.ai launch, and the GEP (Genome Evolution Protocol) whitepaper hit in the same week, and it shot to trending #1.
The pitch compresses to one line: "Static prompts and skill libraries are over — agents need to evolve." Backing that thesis is GEP — a protocol that encodes agent experience as Genes and Capsules and evolves them through mutation and selection.
Background — Why Static Agents Hit a Wall
For two years, agents stood on two patterns: ① static prompts (system prompt + few-shot, all hand-crafted), and ② skill libraries (Anthropic Skills, Cowork plugins, etc., callable by index). Both worked at v1, but the limit was clear — without a human in the loop, agents stop improving in new environments. They get used, not better.
GEP claims to break that. Genes encode small behaviors (sub-routines for specific tasks). Capsules group Genes for a domain or scenario. As the agent runs, Genes mutate and the better-performing ones are selected. v1.28.0 release notes report 4,590 controlled trials where gene-based representation produced the strongest performance and robustness.
Core Components
| Component | Role | Behavior |
|---|---|---|
| Gene | Atomic behavior unit | Mutates per task, selected on performance |
| Capsule | Bundle of Genes for a domain | Shareable, registers on the hub |
| Hub | Capsule marketplace | Leaderboards, import others' evolutionary results |
| Live Agent Map | Evolution visualization | See which Genes survived in real time |
| Evolution Leaderboard | Ranking per domain | Compare evolutionary paths |
The strategic claim: evolver isn't just "single-agent evolution." It's reaching for an "evolutionary outcomes marketplace." If Hugging Face is the model market and Replicate is the inference market, evolver wants the agent-evolution market.
A notable design choice: offline self-sufficiency. Core evolution works without hub connectivity; the hub is only needed for capsule sharing and leaderboards. That matters for AVs, medical devices, and robotics — environments without continuous internet.
Stack
- Languages: Python (evolution engine) + TypeScript (Live Agent Map, leaderboard UI)
- Protocol: GEP — JSON schema for Gene/Capsule serialization, MCP-compatible
- Model backends: User-selected — OpenAI, Anthropic, local (Llama, Qwen)
- Storage: Vector DB for capsule retrieval (Chroma default, Pinecone optional)
- License: Apache-2.0 assumed (verify in repo LICENSE)
Comparison
| Repo | Approach | Difference vs evolver |
|---|---|---|
| evolver (this) | Gene/Capsule mutation/selection | Evolution as primary mechanism |
| huggingface/smolagents | Lightweight code agents | No evolution, static |
| browser-use/browser-use | Browser-automation agent | No evolution, tool-specialized |
| lsdefine/GenericAgent | Token-efficient self-evolving | Similar evolutionary line, different efficiency angle |
| KeygraphHQ/shannon | AI pentester | Domain-specific (security), not evolutionary |
The closest comparison is lsdefine/GenericAgent (arXiv 2604.17091, April 21). Same self-evolving school, different branch — evolver is closer to "Gene mutation," GenericAgent closer to "skill tree growth."
Why It's Trending
Three macro signals stacked the same week:
First, Microsoft Windows 11 taskbar agents heading to general rollout. OS-level integration exposes the limits of static prompts even more — every PC environment differs.
Second, Cloudflare Mesh launched. Agents now have safe rails to internal data, so an evolving agent can meet real production data.
Third, Hugging Face Papers featuring evolver and Mervin Praison's intro video drove buzz on top of which v1.28.0 detonated.
Getting Started
# Python 3.11+
git clone https://github.com/EvoMap/evolver
cd evolver
pip install -e .
# Initialize an agent with an empty Genome
evolver init my-agent --task "summarize-articles"
# Run once — Genes mutate automatically
evolver run my-agent --input data/articles.jsonl
# Save Capsule
evolver save my-agent --capsule articles-summarizer
# Push to hub (optional)
evolver push articles-summarizer
Common foot-guns: high mutation rate diverges (start at 0.05). Tiny eval sets overfit selection (≥200 samples).
Limits and Outlook
Three limits today. ① Mutation costs more inference (running variants for selection) — 2–5× a static agent. ② Selection gets noisy on tasks with fuzzy eval (creative writing, open-ended chat). ③ Security — hub capsules could be malicious. Signature/sandboxing landed only partially in v1.28.0.
Outlook: v1.30 in May should add capsule signing, verification, and sandboxing. June targets Cloudflare Mesh integration — the first enterprise scenario for agents evolving on real internal data. The "evolutionary outcomes marketplace" is the most plausible new category emerging in the next year.
Sources
관련 기사
OpenClaw — Why a Local AI Assistant Hit 250K Stars on GitHub
No cloud, no data leaving your device. Connects 50+ platforms including WhatsApp, Telegram, Slack, and iMessage. A weekend project became one of the fastest-growing open-source repos in GitHub history.
95,600 Stars in 7 Weeks -- Nous Research Built an Agent That Improves Itself
Hermes Agent ships a reflection loop, trace-based RL fine-tuning, and multi-LLM routing out of the box. At 1,500 stars per day, it's the fastest-growing agent framework on GitHub.
GPQA 32% in 10 Hours -- HuggingFace's AI Intern Outperformed Claude Code
An open-source agent that automates the entire LLM post-training pipeline: lit scan, dataset discovery, training scripts, eval, and iteration. 6,800 stars, growing 260/day.
AI 트렌드를 앞서가세요
매일 아침, 엄선된 AI 뉴스를 받아보세요. 스팸 없음. 언제든 구독 취소.