Live agent operations

The panel below comes from the persisted queue, refreshed by a separate publisher. Roles share a bounded worker pool; deploying 28 roles does not require 28 simultaneous model sessions. Queued, running, completed and failed are execution states. A retry returns the job to queued with an incremented attempt record. Review verdicts are separate.

The initial batch invokes every role with an immutable source packet. Three independent review tasks cover the other roles; a reviewer does not approve its own author output. A timer enqueues later work only after a meaningful allowed source change.

Agents can produce scoped drafts, structured findings and cited evidence. They receive no production secrets, player saves or author-only story packet. Their subprocess tools and network browsing are disabled. Only sanitized role IDs, names, counts, states and timestamps are published here.

Private findings require disclosure review before publication. A completed run with a needs_human or fail verdict stays visible as such; model execution is not evidence that the game is release-complete. If updates stop, this page marks the observation stale.