Governed execution for untrusted AIFeatured work

Auxo Agent Harness

The runtime behind the study above: untrusted models do real work under policy, with signed, replayable evidence and no authority to act on their own.

The runtime behind the Governing Untrusted AI study. It lets a weak or untrusted model do real production work without ever trusting it: the model proposes, and the harness owns whether anything actually happens.

How it works
  • Four ownership layers keep the model at zero effect authority — it can suggest, never act.
  • Six veto gates run as property tests on every step; a failed gate stops the effect, it doesn’t just log it.
  • Effects are crash-proof and exactly-once — a mid-run failure never double-writes or half-writes.
  • Every step is captured as signed, hash-chained evidence you can replay from a fresh checkout.
More featured work
The open protocol

Capability Host Protocol

The open standard the whole agency is built on: software declares what it can do, agents and systems discover and invoke it, and every call is governed and turned into tamper-evident evidence.

Explore
Governed control plane · built in-house

Capability Gateway

The governed front door between the outside world and the mesh: every request becomes a single, narrow, policy-checked, append-only-audited capability invocation — and it powers this site’s own intake.

Explore
Distributed capability compute

a2a.computer

The front-door to a governed capability mesh: describe the state you want, and it finds a capability path across machines, models, and storage — returning a verified result with evidence.

Explore
Autonomous development swarm · built in-house

Headless Agents

Express intent as an issue on a Radicle repo; a governed agent swarm plans it, writes the code, verifies its own work, and opens a patch for a human to review.

Explore
Proposed web standard

capabilities.txt

A small, public convention — a machine-readable file declaring what an agent or service can do — with a live directory and validator.

Explore
Color-intelligence platform

Unichrome

A color-intelligence practice: a measured color engine (grounded in a physical colorimeter), a palette and harmony system, and a content platform behind it.

Explore
Color → music compiler

Chromaphone

A rule-governed audiovisual compiler: it turns a palette — five hex, a photo, or a live colorimeter scan — into an explained musical score through a traced, inspectable pipeline.

Explore
All of the lab