Recorded dashboard footage with new narration in one consistent synthetic male voice. The guide introduces himself once. Captions and this transcript follow the new narration; the conversation visible inside the recorded dashboard is the original interaction. Waiting time is removed and private account details are redacted.
00:00:00 — Good day. I'm POE, your guide to Agent Control. This dashboard turns requests into governed work, with a plan, an approval boundary and a result you can inspect. We begin with the job catalogue. Six jobs are registered here. System observation is ready, while the five FreeToken jobs have no eligible worker. Readiness is shown before any work begins.
00:00:25 — A Work Parcel keeps the whole request together: its plan, stages, constraints, routing and evidence. The parent coordinates the work, starting each permitted stage only after its dependencies succeed. Run identifiers and artifacts remain attached to the record. If a required gate fails, dependent stages stay blocked.
00:00:45 — Lanes show where work is assigned and the contract governing its execution. You can inspect the current task, selected model route and continuation baton in one place. In this recording, the qualification lane is idle and waiting for a task. Remote access through the agent protocol is explicitly disabled.
00:01:05 — Sessions connect the interface to governed execution. An authorized operator can create a session, resume it, or cancel it. Agent Control retains the authority and evidence for those actions. The local agent connection is available here. The remote connection is disabled, so the dashboard reports that limit clearly instead of implying it is ready.
00:01:25 — The Systems view brings together registered machines, transports and providers. Each entry has its own observed readiness, so a configured connection is not automatically treated as healthy. The local controllers and local agent transport are available here. The remote transport is disabled, and the provider's separate health probe remains unknown.
00:01:45 — The model registry separates a configured model from a route that has actually been qualified. Here, the selected Codex route supports status explanations and deeper reasoning. Other providers are not assumed available. The interface also distinguishes a working reasoning route from the provider's separate health check, which has not established readiness.
00:02:05 — Routing follows qualified roles and recorded evidence. If a quality gate fails, the system records why and identifies the permitted destination before a model handover. That makes the transition inspectable, while preserving the request and its execution boundaries.
00:02:20 — The Crew gives operational roles a visible face. Cadence coordinates lanes, Quill reviews work parcels, Relay follows tools and execution, Lumen scouts model routes, Rook guards resources, and Verity inspects verification evidence. Their activity comes from recorded runtime events. These characters represent accountable roles; they do not imply six additional autonomous agents. Here, Relay and Verity show completed work while the other roles are idle.
00:02:50 — The Approvals view is where proposed actions wait for the operator. A job start, cancellation or frozen benchmark plan must be reviewed and explicitly approved as the exact sealed request. Reading status, inspecting evidence and asking questions are available without that approval. Schedules and publication have their own controls, and a material change requires a fresh decision.
00:03:14 — Schedules show real registered work and its next execution time. The daily Collingham Facebook collection runs at nine o'clock, UK time. It collects candidates for review; publishing an event remains a separate decision.
00:03:27 — A sealed baton carries the evidence, constraints and continuation state needed by the next stage. Its source, destination, digest and receipt preserve the chain of responsibility. You can inspect what was handed over without exposing hidden model reasoning, and missing records remain clearly unavailable.
00:03:48 — Finishing execution is only part of the story. Verification checks the required evidence before the system claims success. This view lets you inspect that boundary and see which checks passed. The runtime performs the work and its verification, while the conversational guide explains the recorded outcome and makes missing measurements clear.
00:04:10 — Voice has separate parts: OmniVoice generates speech, and Whisper handles recognition. Browser playback begins after synthesis and content checking finish. This adapter does not stream speech, and readiness in a separate messaging channel must be checked independently.
00:04:24 — The Pixel workflow discovers Facebook event candidates and passes them for review. It is scheduled daily and prevents overlapping runs. In the recorded attempt, the observer could not find its configured device, so discovery failed. No events were published. The failure stays visible, and publication remains a separate governed step.
00:04:45 — The Transcript control opens the retained conversation through the authenticated interface. It keeps the initiating request and subsequent responses together, giving the operator a readable record to inspect or export.
00:04:58 — Here, the operator explicitly approves the sealed request.
00:05:02 — Relay records evidence from the initial observation and seals it for the dependent stage. Verity records the receipt. The next observation continues from that preserved evidence, with the handover attached to the work.
00:05:15 — Now the requested set is complete: four jobs were requested, and all four succeeded. None failed, remained blocked or awaited verification. No work is still active. This statement follows the actual terminal records. A completed job does not itself imply publication; its declared effects and evidence determine what happened.
00:05:35 — The selected parcel shows its specific objective: two read-only system observations, connected in sequence by a baton and receipt. Every planned job passed the normal verification boundary. Because this was a deterministic control action, no model invocation is claimed, and unavailable provider totals remain unreported.
00:05:53 — The stop control interrupts speech while executing jobs continue.
00:05:58 — The final regression passed all one thousand eighty-three tests, with no failures or skips. The recorded run took one hundred and thirty-five seconds. The release and documentation links below provide the details.