Colophon
Build fast. Ship authorized.
An agent mesh for engineering teams. Specialist agents resolve work peer-to-peer within explicit trust bounds. When agents disagree, an adversarial reviewer escalates to a change control board — where agents and humans decide together. Every decision lands in an immutable audit trail. Humans stay in charge, but only on the calls that matter.
Read more
Agent mesh
Specialist agents collaborate peer-to-peer within explicit trust bounds. When peer resolution succeeds, the work moves. When it fails, the disagreement moves upward through the mesh — carried by an adversarial reviewer — until it reaches a decision.
A three-tier diagram showing how Colophon's specialist agents resolve work peer-to-peer at the bottom tier, escalate to arbiter agents in the middle tier only when specific conditions are met (a flag is raised, an audit sample triggers, or peer resolution is exhausted), and reach the human only when arbiters themselves escalate. Visual asymmetry: the peer tier is wide and prominent, the arbiter tier is narrower, the human tier is a single node.
Chain of custody
Every work unit passes through four review lenses — code, adversarial, test quality, security — before it ships. Overrides exist and are tracked. Nothing bypasses the audit trail.
A horizontal pipeline showing how work moves through the Colophon chain of custody. Stages from left to right: requirements, trace review, architecture decisions (optional branch shown above the main path), decomposition into work units, implementation with attached test evidence, the review gauntlet with four parallel review lenses (code, adversarial, test quality, security), and release. Each transition between stages is a structural gate enforced by the chain — not convention. The implementer can refuse to start work whose upstream is not ready, and the refusal is the protection.
A vertical stack of five verification layers. The top four layers are the four review lenses applied to every work unit — code, adversarial, test quality, and security — in the order they run as part of the release gate. The bottom layer is live-endpoint verification, which runs against the deployed site after a release to confirm the production path is serving what the chain of custody shipped.
Four review lenses gate every work unit before release. A fifth layer — live-endpoint verification — runs against the deployed site after a release. Every layer stays in the chain; nothing bypasses the audit trail.
Agent team
A generalist team covers planning, requirements, architecture, implementation, the four review lenses, release coordination, documentation, and root-cause analysis. Specialist agents add domain depth — iOS, Android, Hono.js, AWS, Postgres, DynamoDB, web frontend, authentication, observability, and design system — and inherit the same chain of custody. Adding a specialty is adding an agent, not rebuilding the mesh.
Change control board
Immutable requirements traceability, end to end. Every requirement becomes work units. Every work unit passes review lenses. Every review lands in the decision log. Every release cites the chain that produced it. Agents and humans decide together on the board. Drop in at any point and read back to the original decision.
Automated state
The Gantt and chain graph update as work moves. Every release re-renders the state of the mesh. No manual refresh, no stale dashboards, no hand-maintained reports.
When something breaks
A root-cause agent traces the failure back through the audit trail to the decision that produced it. No guessing. No post-mortem by recollection.