The namespace itself: every source publishes on automap/v1/<enterprise>/<site>/<area>/<line>/<asset>/<branch>/<signal>, agents at level 3 under operations/agents. Counts and the worst quality flag roll up to every level. Click a signal to inspect its last envelope.

The edge gateway tags every message with a quality flag (good / uncertain / bad / stale), buffers when the network drops and backfills afterwards. Fault injection below proves that the flags, store-and-forward and the agents' abstention work. Nothing here touches a real network.

Read-only SQL (ENG / TEC). One SELECT, capped at 500 rows. Tables: genealogy (parts, ann_predictions, qc_outcomes, prediction_scores, material_cards, work_orders, kpis …) or historian (messages, telemetry_1min).

Every agent on the platform, checked live and editable, at the level of detail you choose: 1 Glance one line - its checks, its grade, its last word; 2 How it works what it hears, how it reasons, what it says; 3 Wiring every input with who publishes it, every output with who listens; 4 Settings every setting with its default, what it does and when a change takes effect - ENG edits, gives a reason, saves; 5 Evidence the checks, the score against its baseline, what it said lately and its code. The buttons above set the level for all; each agent's own 1-5 opens just that one. Edits are kept for this site, logged with who and why, and published on the agent's own topic. Build new agents in New agent.

Versions are never deleted; promote and rollback move the active pointer. A campaign CSV or a JSON model bundle from AutoDT-AutoForm enters as a candidate and is promoted only if it beats the active model on this plant's own prediction-vs-reality pool.

Derived from the live contracts: an edge means a subscription pattern overlaps a publication. Agents already work together through the namespace - one agent's prediction is another's input.

What each role sees. The tabs only choose the screen you look at; the header role (now ENG) decides what you may do. Every suggestion is explainable and overridable; a command is only ever applied by a human-confirmed button.

A decision agent takes other agents' outputs as its inputs, matches them per part in whatever order they arrive, and combines them into one advisory decision - RELEASE, REVIEW or HOLD - for a human screen: by ordered rules (first match wins) or by a weighted vote. Dry run replays the recent traffic through it before you create it. It scores itself against the QC outcome of the same part. A learned combination is a learned agent whose inputs include the Agents feature group.