Blog

Notes on the tools we build and how we build them.

Why we built Concord: organized product development at chaotic speed

Several AI coding agents can move quickly on one machine. The hard part is keeping their work aligned when each change can look correct by itself. Concord gives that work one durable record and one set of admissible operations.

A durable record for the work

Concord stores approved requirements, work state, evidence, knowledge, and research in a local SQLite database. Agents reach that state through typed, authorized operations, so the engine can refuse an operation that does not satisfy the current boundary.

Completion must carry proof

The operator approves outcome predicates before work starts. At acceptance, Concord compares the delivered result with those predicates and requires verification, review, and commit evidence before a work item can finish.

Conflicts stop before they spread

Each change declares the architectural Domains it writes. When two changes claim a conflicting Domain, Concord refuses the second change until the overlap has a recorded resolution. The system names the failed condition and its remedy instead of asking an agent to guess.

Concord is an MIT-licensed Go core with SQLite as its sole durable authority. Its OpenCode adapter carries session context and typed transport, while domain logic and recovery stay in the core.