AGENTS

Software, described like software. Each of these has a job, a tool grant, a way of being checked, and a set of things it reliably gets wrong — and the last one is the field that took the longest to write and is worth the most.

Names, routing descriptions, tool grants and model tiers on this page are read from the definition files themselves by pnpm agents:sync, and a test re-derives them on every run. The registry cannot claim a tool the harness never granted.

REGISTRY · 7

Every agent in the lab, with its model tier, tool count, and the project it belongs to.
StatusNameModelTools
ACTIVEclaims-auditorChecks whether the site is still telling the truth about itself, across every surface that repeats a claim.opus8
ACTIVEdoubtAudits the implementation against its own claims, with no write access at all.opus10
ACTIVEperf-sentinelMeasures whether a change moved performance, using an alternating A/B rather than a before-and-after.sonnet4
ACTIVEplannerTurns a scout report into file operations exact enough that another agent can execute them without judgment.opus6
ACTIVEremediatorFixes what the audit found, and is forbidden from making a finding disappear by editing the check.opus6
ACTIVEscoutReads the codebase before anything is written, and says out loud whether the premise is already stale.sonnet8
ACTIVEship-captainRuns the gate sequence, reconciles with upstream, ships, and then verifies the deployment rather than the build.sonnet4

7 agents · 46 tool grants · 4 of them cannot write anything at all

WHY THE READ-ONLY ONES

4 of these 7 cannot write a byte — claims-auditor, doubt, planner, scout — and that is the design rather than an oversight. Given a failing check and a pen, the cheapest available repair is always to change the check, and an agent will find that repair every time. Splitting the auditor from the fixer costs an extra pass and buys the one thing a single agent cannot: a finding that cannot be resolved by rewriting the evidence.

The same reasoning runs the other way for remediator, which does hold a pen and is forbidden from touching a test, a fixture, a snapshot or an assertion in order to make a finding go away. A check that genuinely looks wrong is a new finding to raise, not one to resolve.