#Atlas and Argus

Two agents run the pipeline. The builder agent drafts; the verifier agent reviews. In the UI they appear as Atlas and Argus: as artifact owners, work-item assignees, and in the Agents section of a project's team roster.

#The builder agent

The builder agent drafts artifacts, generates documents, writes code, and opens pull requests. Artifacts it creates show Owner: Atlas. When the assistant carries out a task in a sandbox, the builder does the work, operating as the signed-in user's delegate with your confirmation at every write.

The builder can also be asked to investigate a reported issue and prepare an implementation plan for your approval before any code changes.

#The verifier agent

The verifier agent reviews every draft before a person reads it. It produces findings against a pass threshold and runs up to four revision cycles if the draft does not clear it. You see its findings on an artifact under the Argus panel.

When all findings resolve and the score passes the threshold, the artifact moves toward human approval automatically. When the draft cannot converge in four cycles, the item pauses and surfaces as a pending action.

You can dismiss a finding by supplying a rationale, or steer it by supplying guidance that tells the builder how to address it. Both are recorded.

#Model routing

Every model call names a work-type bucket. The model behind each bucket is resolved from the platform configuration, overridden at the project level, with baked defaults as the final fallback.

Project-level settings override the platform defaults for that project's pipeline. See AI providers and models for how to configure each bucket.

#Next steps

• AI providers and models: configure model routing per bucket.
• Audited actions: how the builder's writes are recorded.
• Pending actions: when a verifier cycle cannot converge, it surfaces here.