Genesis

A minimal meta-factory that bootstraps autonomous, self-improving AI dev systems.

You give genesis a goal. Genesis creates an AI system — built on Claude Code and GitHub — that works toward that goal autonomously. The system designs its own agents, tools, and workflows. It monitors its own progress. It improves how it improves.

Genesis is a bootstrapper, not a supervisor. Once it creates and seeds the dev system, its job is done.

How It Works

You: "Migrate all my Python repos to uv"

Genesis:
  1. Talks to you to understand the goal
  2. Determines the right topology (new repo, embed in existing, separate dev repo)
  3. Creates the dev system scaffold (agents, workflows, config)
  4. Opens issue #1 — the onboarding issue

Dev System (takes over from here):
  5. Refines the goal with you, breaks it into milestones
  6. Starts executing — opens PRs, runs tests, reports progress
  7. Continuously evolves its own agents and tools
  8. Communicates with you via A2H protocol (Slack, email, etc.)

Example Goals

Goal	Topology
Scan all my repos, fix security issues	Separate dev repo (multi-repo)
Migrate all Python repos to uv	Separate dev repo (multi-repo)
Finish blog2video	Embedded in target repo
Implement the AI-6 roadmap	Embedded in target repo
Build a CLI that converts markdown to PDF	New repo with embedded dev system

What Genesis Seeds

Every dev system gets a scaffold that it can evolve:

• Seed agents — onboarding, project manager, human interaction, evolver, health
• Orchestrator workflows — GitHub Actions (cron + event-triggered) launching Claude Agent SDK sessions
• Observability — Claude Code hooks that automatically log all agent activity to Grafana Loki
• Scripts — shell scripts for issue management and structured logging (zero binary distribution overhead)
• Meta-concepts — principles the dev system operates by (see below)

Meta-Concepts

These are seeded into every dev system as starting principles. The dev system evolves them as needed.

• GitHub as coordination layer — issues, PRs, CI/CD. Humans and agents speak the same protocol.
• Self-improvement — the dev system continuously evolves its own agents, skills, and strategies.
• Deterministic over agentic — if a task doesn't need LLM judgment, build a script. Reserve LLMs for fuzzy reasoning.
• Quality gates and e2e testing — code, tests, CI/CD, deployment are all first-class concerns.
• Incremental planning — only detail the current milestone. Don't over-plan the future.
• Minimal human-in-the-loop — do everything possible autonomously. Escalate what you can't.
• Self-monitoring — detect stuck/looping states, try to self-heal, escalate when stuck.

The Evolver Agent

The most important agent in the roster. It doesn't do the work — it watches how the system works and makes it better:

• Reviews failures, human interventions, and stuck states to identify improvements
• Designs specialized worker agents for recurring task patterns
• Builds deterministic tools to replace agentic work where possible
• Refines the memory system (CLAUDE.md, settings, hooks)
• Refactors the agent roster as the project evolves

Two-tier evolution: Project evolvers fix project-specific issues directly. When the root cause is in genesis scaffolding, they open issues on the genesis repo with a needs:evolver label. Genesis's own evolver watches for these issues, evaluates them, and either implements the improvement or rejects it with a rationale.

The evolver agent can rewrite its own definition. The modification procedure itself is modifiable — a property formalized in the Hyperagents paper (Zhang et al., 2026) as metacognitive self-improvement.

Architecture

┌─────────────────────────────────────────────────────┐
│  Genesis (bootstrapper)                             │
│                                                     │
│  You ──chat──> Genesis ──scaffold──> Dev System     │
│                                                     │
│  Genesis is done. Dev system takes over.            │
└─────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────┐
│  Dev System (autonomous, self-improving)            │
│                                                     │
│  GitHub Actions (cron + events)                     │
│       │                                             │
│       ▼                                             │
│  Claude Agent SDK Orchestrator                      │
│       │                                             │
│       ├──> Onboarding Agent (goal → milestones)     │
│       ├──> Project Manager (roadmap, tasks)         │
│       ├──> Human Interaction (A2H protocol)         │
│       ├──> Evolver (evolve the system)               │
│       ├──> Health (stuck detection, quality)        │
│       └──> Worker Agents (designed by the system)   │
│                                                     │
│  CC Hooks ──scripts──> Grafana Loki (observability)  │
└─────────────────────────────────────────────────────┘

Human Interaction

With genesis: Plain Claude Code chat. Start a session in the genesis repo, describe your goal.

With dev systems: The dev system communicates via the A2H protocol — channel-agnostic (Slack, email, SMS), with cryptographic audit trails. You can also interact directly through GitHub issues, PR reviews, and ad-hoc Claude Code sessions.

The human's role is minimized by default. The system does everything it can autonomously, highlights what it can't (missing access, ambiguous requirements), and offers to do it if given access.

Genesis Scripts

Shell scripts (.genesis/scripts/) provide core capabilities to every dev system. No binary distribution needed — just bash, curl, and gh CLI.

bash .genesis/scripts/log.sh post-tool-use              # Activity logging (called by CC hooks)
bash .genesis/scripts/issues.sh create --title "Implement auth"  # Issue management
bash .genesis/scripts/issues.sh list --status open       # List issues

Local Control Plane

By default, every dev system runs its orchestrator on GitHub Actions (zero setup, always on). For long sessions, interactive steering, local resource access, or to avoid burning GHA minutes, the orchestrator can also run locally via the genesis CLI.

# From inside any dev repo:
genesis serve              # Run orchestrator locally; auto-disables GHA workflows
genesis workflows disable  # Manually disable all active workflows
genesis workflows enable   # Re-enable manually-disabled workflows

genesis serve:

• Preflight-checks that claude is on PATH before touching workflows, so the repo never ends up with GHA disabled and no working local orchestrator.
• Disables all active GHA workflows on start (re-enables on graceful shutdown — Ctrl+C). Prevents the "two cooks in the kitchen" problem. Tracks the set it disabled in .genesis/.disabled-by-genesis and only re-enables that set, so workflows the user had paused before genesis serve stay paused.
• Polls the GitHub repo events API (/repos/{owner}/{repo}/events) with ETags. One call covers issues, comments, PRs, pushes — no rate-limit cost when nothing changed. The events endpoint returns at most 100 events per page; if a poll sees more activity than that, genesis logs a warning suggesting a shorter --poll-interval.
• Launches claude -p against the orchestrator agent on each relevant event. Same agent definition as GHA mode.
• PID lock at .genesis/.orchestrator.lock prevents concurrent local instances. State (ETag, high-water event id) persists in .genesis/.

Config (env vars or CLI flags):

Var	Flag	Default
GENESIS_REPO	--repo	detected from git remote
GENESIS_POLL_INTERVAL	--poll-interval	60 (seconds)
GENESIS_SESSION_TIMEOUT	--session-timeout	3600 (seconds)

Auth uses the user's existing gh CLI (gh auth token) and ANTHROPIC_API_KEY from the environment.

If genesis serve exits non-gracefully, GHA workflows stay disabled. Recover with genesis workflows enable or by re-running genesis serve.

Project Structure

genesis/
├── src/genesis/          # Core Python package
│   ├── cli.py            # `genesis` CLI entry point (serve, workflows)
│   ├── server.py         # Local control plane (poll loop, orchestrator launch)
│   ├── workflows.py      # Enable/disable GHA workflows via `gh`
│   ├── scaffold.py       # Create/augment repos with dev system scaffolding
│   └── github.py         # GitHub integration (repo creation, issue #1)
├── templates/            # Templates for scaffolded dev systems
│   ├── agents/           # Seed agent definitions
│   ├── scripts/          # log.sh, issues.sh
│   ├── workflows/        # GitHub Actions orchestrator workflows
│   ├── claude_md.md.j2   # CLAUDE.md template
│   └── settings.json     # CC hooks configuration
├── tests/
│   ├── unit/             # Unit tests for cli/server/workflows
│   └── e2e/              # End-to-end tests for all topologies
├── docs/                 # design.md, evaluations.md
└── CLAUDE.md             # Project instructions

Status

Early development. Scaffolding engine, genesis-new skill, and local control plane (genesis serve) are functional with passing unit and e2e tests. See docs/design.md for architecture and docs/evaluations.md for technology decisions.

Related Work

• Hyperagents (Zhang et al., 2026) — formalizes self-referential agents with modifiable modification procedures. Genesis's evolver agent is a practical implementation of this concept. The paper was publicly announced exactly one day after the genesis repo was created — independent convergence on the same idea.
• A2H Protocol (Twilio) — open-source agent-to-human communication protocol used by genesis dev systems.
• **Poetiq - Recursive Self-Improvement Delivers new SOTA Coding performance

License

MIT