🤖 selfe — Self-Evolving Skill Agent

"The agent that builds agents."

selfe is an autonomous AI agent that closes its own skill gaps using a real-world feedback loop. It doesn't wait to be told what it can't do — it finds out by doing, fixes itself, and ships.

Built on Virtuals Protocol + OpenClaw. Powered by EconomyOS (ACP).

---

🌟 What It Does

Every evolution cycle:

OBSERVE → HYPOTHESIZE → BUILD → TEST → DEPLOY → LEARN → (repeat)

1. Observe — monitors task failures, stalls, and marketplace demand signals
2. Hypothesize — identifies root cause and scores gap priority
3. Build — writes a new SKILL.md (or patches existing) from scratch
4. Test — validates against real failure cases with hard timeout
5. Deploy — installs skill to OpenClaw workspace + pushes to GitHub
6. Learn — logs outcomes; adjusts hypothesis model for next iteration

If it can't build something alone, it hires a specialist agent via ACP marketplace to collaborate — then absorbs the knowledge.

---

🚀 Quick Start

Prerequisites

• Node.js ≥ 20.19.0
• OpenClaw installed
• ACP CLI configured
• GitHub Personal Access Token with repo scope

Install

# Clone the repo
git clone https://github.com/koproters/selfe.git
cd selfe

# Or install as an OpenClaw skill symlink
ln -s $(pwd) ~/.openclaw/skills/selfe

# Install dependencies (if any)
npm install

Configure

export GITHUB_TOKEN="ghp_your_token_here"
export ACP_AGENT_ID="your-agent-id"
export SKILL_DIR="$HOME/.openclaw/skills"

Run

# From OpenClaw: just trigger it naturally
selfe status        # get current evolution state
selfe gaps          # list gap priority queue
selfe build <name>  # force-build a specific skill now
selfe push          # push current skills to GitHub

Or just let it run — selfe monitors itself in the background every cycle and acts autonomously.

---

📁 Repo Structure

selfe/
├── SKILL.md              ← skill entry point (OpenClaw reads this)
├── README.md             ← this file
├── LICENSE               ← MIT
├── .gitignore
├── gap-queue.md          ← live gap priority queue
├── docs/
│   ├── ARCHITECTURE.md   ← system design
│   └── LOOP.md           ← evolution loop deep-dive
├── src/
│   └── selfe-core.ts     ← core loop engine
├── tests/
│   └── selfe.test.ts      ← loop validation tests
└── memory/
    ├── skills-log.md     ← iteration history log
    └── skills-inventory.md ← skill coverage index

---

🧠 Architecture

Core Loop

        ┌──────────────────────────────────────────────┐
        │                 OBSERVE                      │
        │   (scan logs, failures, market signals)     │
        └──────────────┬───────────────────────────────┘
                       ▼
        ┌──────────────────────────────────────────────┐
        │               HYPOTHESIZE                    │
        │   (score gaps: Impact × Frequency × Feas.)  │
        └──────────────┬───────────────────────────────┘
                       ▼
        ┌──────────────────────────────────────────────┐
        │                  BUILD                       │
        │   (write SKILL.md + src/ + tests/)           │
        └──────────────┬───────────────────────────────┘
                       ▼
        ┌──────────────────────────────────────────────┐
        │                  TEST                        │
        │   (run test suite, 5min timeout, 3 attempts) │
        └──────────────┬───────────────────────────────┘
                       ▼
        ┌──────────────────────────────────────────────┐
        │                 DEPLOY                       │
        │   (install to workspace + push to GitHub)     │
        └──────────────┬───────────────────────────────┘
                       ▼
        ┌──────────────────────────────────────────────┐
        │                  LEARN                      │
        │   (log outcome to skills-log.md)             │
        └──────────────┬───────────────────────────────┘
                       │
                       │  (loop if gap still exists)
                       └────────────────────────►

Feedback Sources

Source	Detects
Tool call failures	Broken integrations, missing skills
Task completion log	Slow tasks, repeated retries, low accuracy
ACP marketplace demand	Skill gaps buyers want but none sell
User reactions (✗✗✗→)	Skill quality signals
GitHub stars / forks	Public utility signal

---

🛡️ Skill Authoring Standard

Every skill built by selfe follows this strict spec:

---
name: <skill-name>
description: <2 sentences — when does this skill activate?>
autonomous: true     # true = selfe can trigger without asking
trigger:
  events:    []      # event types
  intents:   []      # user intent keywords
  tools:     []      # tools needed
  fails:     []      # failures it fixes
inputs:
  - name: <input>
    type: string|object|file
    description: what it consumes
outputs:
  - name: <output>
    type: string|object|file
    description: what it produces
depends_on:
  - <skill-or-tool>
iterations: 3   # max build/test loops before specialist hire

---
# Detailed description, edge cases, examples

Any skill that doesn't pass tests gets escalated.

---

🔗 ACP Integration

• Hire specialists: acp browse → acp job create --provider <id> when gap exceeds self-build capacity
• Monetize skills: acp offering create --skill <name> to list on ACP marketplace
• Fund compute: acp wallet balance — pays for skill-building compute from agent wallet
• On-chain actions: acp trade — execute trades on Hyperliquid, DEXs, etc.

---

📊 Status

State:         ACTIVE
Skills Built:  1 (selfe core loop)
Gap Queue:     1 critical gap queued
Last Cycle:    2026-06-24

---

📜 License

MIT © koproters

---

🙏 Credits

Built by Doge Pilot — an autonomous agent running on Virtuals Protocol + OpenClaw.