Run AI coding agents like a Scrum team.
Talk to an agent about what you want built — it files the work, a hybrid AI team delivers it, you watch every step, and it ships as a pull request. A cloud model is your Tech Lead, local (or cloud) LLMs are the dev team, and you stay the Product Owner — two human gates and hybrid cost routing built in.
AI agents write code well. They're bad at being trusted.
Today's coding agents are opaque autopilots: you approve one giant diff at the end, every token runs through an expensive frontier model, and on a big repo they drift. There's no agile structure, no per-step checkpoint, no cost control.
Two human gates
You approve the decomposed backlog before any code runs, and accept finished work at the end. Everything between is automatic but visible.
Hybrid cost routing
Work stays local and free until a retry budget is exhausted, then escalates to the cloud. You don't pay frontier prices to read every diff.
Atomic ticket contracts
Each ticket is a JSON-Schema contract with code injected inline and a machine-verifiable Definition of Done. "Done" is a test result, not an opinion.
Roles, assigned where they belong
The non-obvious choice: the human is the Product Owner, not the Scrum Master. Process mechanics get automated. Product judgment doesn't.
From a conversation to a pull request
Not another coding agent — the control plane
| HAAO | Coding agents (Devin / Cursor / Copilot Workspace) | |
|---|---|---|
| Human control | Two explicit gates (scope + acceptance) | Single end-of-run diff approval |
| Cost model | Hybrid: frontier for reasoning, local for execution | Frontier model for everything |
| Privacy | Local execution option (code stays in-house) | Cloud-only |
| Mental model | Agile tickets + Definition of Done | Freeform autopilot |
| Trust posture | Audit verdict + rework loop, reversible | Mostly opaque |
See the whole flow in your browser
An interactive simulation of the full loop — no live model calls, no setup.
Frequently asked questions
What is HAAO?
HAAO (Hybrid AI-Agile Orchestrator) is an open-source governance layer for AI coding agents. A cloud model decomposes and audits, local models execute, and two human gates keep a person in control of scope and what ships.
How is it different from Cursor, Devin, or Copilot Workspace?
Those are coding agents — one model does everything and you approve a diff at the end. HAAO is the control plane around agents: hybrid cost routing (frontier for reasoning, local for execution), two explicit human gates, and tickets as machine-verifiable contracts.
Do I need cloud API access?
The cloud model (Claude) handles decomposition and audit. Execution runs on local models via LM Studio, so the bulk of the work is free and private. Cloud is gated behind a retry budget, not in the hot path.
Is it production-ready?
No — it's a working end-to-end prototype, validated on real open-source repos (see Benchmark: 82% local-finish, 46% one-shot, across small and large files). It's deliberately single-project and single-worker today. Treat it as an early, measured prototype — not production software.
Does it ship code, or just write it?
It ships. When you accept a result, HAAO opens a branch and pull request to your GitHub or GitLab — least-privilege and idempotent (one PR per ticket).
Can I use OpenAI or other cloud models, not just Claude?
Yes. Register multiple cloud providers (Anthropic, OpenAI, OpenRouter and more) with encrypted keys, and assign a specific model per role — or keep everything local.
Is it safe to run AI-written code?
Test/DoD commands run sandboxed with networking disabled and a scoped environment; secrets are encrypted at rest and redacted from logs; injected repo content is treated as untrusted. You can also require an API token.
Is it open source?
Yes, MIT licensed. You can self-host and bring your own keys.