An autonomous software build pipeline that turns a project description into working, tested code — with human approval gates, tamper-evident audit trails, and strict quality contracts that keep AI-generated output on-spec.
Built on LangGraph. Compatible with Claude and OpenAI.
Try it without an API key — hosted demo, 3 runs per session.
When you ask an AI model to write an entire software project, three things consistently go wrong:
- Drift. The AI forgets decisions it made earlier and contradicts itself across files.
- No receipts. You can't prove what the AI was asked, what it produced, or whether anyone reviewed it.
- All-or-nothing. The output is either accepted wholesale or thrown away — there's no structured quality gate.
The engine addresses all three by wrapping AI code generation in a pipeline with formal contracts, evidence-based testing, and a cryptographically signed audit trail.
You describe what to build
↓
Engine creates a binding design contract (JSON blueprint)
↓
AI writes code in chunks, each checked against the contract
↓
Code is extracted into a standalone project folder
↓
Automated tests run: syntax, imports, linting, type safety, contract compliance
↓
A verification report gives a go/no-go recommendation
↓
Every step is recorded in a tamper-evident audit log
At any stage, the pipeline can pause and ask a human to approve before continuing. You stay in control; the AI stays in its lane.
git clone https://github.com/Kaden-G/autonomy_engine.git
cd autonomy_engine
cp .env.example .env # add your ANTHROPIC_API_KEY
docker compose up # → http://localhost:8501Create a project, pick a tier, launch a build. More options (local install, CLI-only): docs/usage.md.
| Topic | Doc |
|---|---|
| Pipeline stages, orchestration, contracts, tiers, gates, project layout | docs/architecture.md |
| Intake, running, dashboard, supported project types, development setup | docs/usage.md |
Env vars, config.yml, models.yml, DECISION_GATES.yml, efficiency, resilience |
docs/configuration.md |
| Security model, prompt-injection defense (OWASP LLM01), POAMs, framework mappings | docs/threat-model.md |
HMAC chain, verify_trace CLI, CI enforcement, key management |
docs/audit-trail.md |
| Migration history (Prefect → LangGraph in v2.0) | docs/migration-langgraph.md |
| Setup, QA, test suite, filing a change | docs/contributing.md |
The engine's threat model is documented at docs/threat-model.md, including explicit non-goals. If you discover a security issue, please open a GitHub issue with the security label — or for sensitive reports, contact the maintainer directly (see commit Author metadata).
The project's HMAC audit trail is verifiable with python -m engine.verify_trace --run-id <id>. CI enforces tamper-detection on every PR via the trace-integrity job.
MIT — see LICENSE.
Kaden-G. This project was built with Claude as a development partner; see docs/contributing.md for the approach rationale.