Agents

A coding agent works better on a repository that gives it: one obvious place for new code, an explicit architecture, generators instead of free-form file creation, executable constraints instead of prose it can skim past, and fast feedback when it gets something wrong. OpenLoam is built so a model doesn’t have to rediscover the application’s tenancy and authorization model from scratch on every task — it reads AGENTS.md once, and the rest is enforced.

business requirement
        ↓
coding agent
        ↓
AGENTS.md — the agent contract
        ↓
generators + conventions
        ↓
implementation
        ↓
guardrails + tests  ──fail──> agent corrects, loop repeats
        │
       pass
        ↓
      done

Start here

  • The agent contract — what AGENTS.md is, what an agent should read first, and the invariants it commits to.
  • Guardrails — how those invariants are enforced as failing tests, not just documentation.
  • Golden tasks — the benchmark methodology, and the real (internal, disclosed) numbers on whether any of this actually changes agent behavior.

Subsystem deep-dives

AGENTS.md is byte-budgeted (≤32 KB), so anything past a one-line summary in its table lives here instead. Read the relevant page before touching that subsystem.

  • Foundation — the tenancy, authorization, and audit guarantees the agent contract assumes.
  • Reference — configuration and generator syntax.