The Sentinel workspace

The Sentinel workspace

Sentinel in the left navigation is the control plane for AI agents in production. Every surface below reads the same data — runs, their evaluations, incidents, monitors — and every list row deep-links to the run, agent, incident or monitor it names, so an investigation never dead-ends in a summary.

Tab	Route	What it is for
Control Plane	/sentinel/overview	The fleet at a glance: the run funnel (received → evaluated → clean → breached), each agent's health, and the change digest — what moved since last time, computed from the data rather than written by a model: quality shifts, new or silent agents, cost moves, safety breaches, monitors firing or blind, connectors failing, judge disagreement.
Runs	/sentinel/runs	Every run, filterable by agent, version, state, score band and time. Open one for its events, per-dimension scores with rationale, the incident if it breached, and the feedback controls.
Explorer	/sentinel/explorer	Group and filter runs by any dimension, pick measures, save the view, export CSV. Every stock dashboard widget drills into it.
Dashboards	/sentinel/dashboards	Six stock dashboards — Fleet Health, Agent Quality, Evaluation Pipeline, AI Cost & Efficiency, Quality Detail, Safety & Automation — plus one-click provisioning of the same views into the platform's dashboard list.
Sentinel Health	/sentinel/health	Sentinel watching itself: evaluation backlog and its age, evaluations in flight, failed and partial, the judge's own token cost, monitors firing, recovery deliveries failing, unproven connectors, query latency.
Incidents	/sentinel/incidents	Breaches, each with root-cause analysis and recommendations to approve or dismiss.
Behaviour	/sentinel/patterns	Recurring behaviour patterns learned from evaluations — the same ranking the coaching uses, shown decomposed (recurrence × recency) so the order can be argued with. Resolve a pattern once it is fixed.
Actions	/sentinel/actions	Recovery connectors — where approved actions are delivered — and the delivery log with retries, failures and refusals.
Monitors	/sentinel/monitors	Fleet-level conditions with a preview of how often they would have fired, cooldown, mute and an owner.
Agents	/sentinel/agents	Register agents, get their keys and connection snippets. Open an agent for its own workspace: journey, runs, versions compared, evaluations, incidents, recommendations — never mixed with other agents.
MCP	/sentinel/mcp	The MCP server configuration and tool list, and the audit of every call an agent has made.
Configuration	/sentinel/configuration	Policy and recovery (breach threshold, autonomy level, halt switch, platform alerting), the knowledge base for grounding, and custom evaluator profiles.

Search sits in the header of every tab and finds runs by execution or trace id, agents, versions, campaigns, fleets, environments, judge models, monitors, incidents, connectors, saved views and Explorer dimensions — capped per category so one term matching ten thousand runs cannot bury the one monitor.

Reading a score

Scores are 0–100 per dimension. Colour bands are fixed — ≥ 70 good, 40–69 warning, < 40 poor — so a regression reads the same everywhere. The line that matters is the breach threshold (default 50, set under Configuration, overridable per dimension): a dimension below it breaches the run, opens an incident and counts toward the safety and fleet monitors. A run the judge could not score is shown as skipped with the reason, never as zero.

Cookbook: from a bad answer to a fix

1. Control Plane. The change digest says support-bot quality down 18 points since yesterday. Click it.
2. Runs, filtered to that agent and the poor band. Open one: factual grounding 22, rationale quotes the claim the knowledge base does not support.
3. Incidents. The breach has an RCA naming the version that introduced it and a recommendation — require citations. Approve it; it is delivered through the connector for that agent.
4. Agent → Versions. Confirm the delta between 1.3 and 1.4 is real: the compare states the difference, its direction, and refuses to conclude below ten runs a side.
5. Monitors. Add a version regression monitor so the next one pages before a person notices.

How it relates to Workbooks

Sentinel and the Workbook AI nodes share one evaluator engine and one deterministic safety backstop. The difference is when they run: a Workbook evaluates on a schedule or in CI against inputs you chose; Sentinel evaluates the traffic your agents actually served. Use both — the eval suite catches a regression before deploy, Sentinel catches the one that only appears with real users.