Control

reading the estate…

Monitor deployed workflows, active runs, approvals, performance, cost and control events. Each measure is calculated from platform records and opens to its supporting detail.

Page window · click a tile’s window to override it
Lockdown
Attentionreading the queue…every 5 s
Deployed agents…
Active now…nothing in flight
Awaiting approval…no gates open
Runs…every journal counted
Spend…
Avg run time…wall clock
Guardrail denials…click for the rule that fired
Kills…operator decisions
Agent lifecyclethe latest version of each agent, each stage lit only by its own record: a version, a review, a gate decision, a deployment, a retirement
AgentDesignedReviewedApprovedCanaryDeployedRetired

Reading the registry…

Hover a stage for the record behind it and what a click does.

Estateevery agent, its whole record — click a row, tick rows for a fleet action
AgentStatusBenchmarkHealth · last 14Avg timeSpend · 7d

Reading the registry…

Health: one tick per run — hover for its state and age, click to open.

Recent runsAll runs →
Live nowSIGTERM, journalled

Reading the process table…

Inboxgates waiting for a person · tick several to answer together

Reading the inbox…

Guardrailspolicy in action, 7 days

Every refusal, gate and kill is a journal entry with a because-chain — open any run to read the exact rule that fired.

FinOpsjournalled burn vs budget
—Billed · MTD
—Billed · 7d
$0.00Journalled · 7d
unsetBudget
Azure billing not connected: set the AZURE_* variables
Activity14 days — click a day
costtoday

Hover a day for its runs and cost; click to open them.

Model mix7d — click a model

No model calls in this window.

Trigger queueexternal posts, by agent

The queue needs the registry database.

Cost · accuracy · time7 days — one triangle per model
Accuracyeval pass rateCostcheaper per runTimefaster wall clockThe outline is the ideal. The shape inside is the model. Dashed corner: no evidence on that axis yet.

No model calls in the last 7 days.

Cost and Time bars are log-scaled against the best model in view: full is the best, half is ten times worse, empty is a hundred times worse. Accuracy is the eval pass rate measured on that model. Click a model for the runs behind it.

Artifactswhat the runs wrote
Partner agentsA2A peers, from the registry

Spoken to only once registered and approved, signed both ways. Their answers arrive marked untrusted — the same taint rules as every other boundary.

Least privilegegrants × declared risk × gates
Execution ledgerappend-only, exportable
…Entries
…Runs
…Last write

Every model call, tool call, refusal, gate and kill — one entry each, carrying the evidence chain it rests on. The journal is what an auditor reads; nothing on this page exists that it doesn’t.

Download NDJSON

reading the estate…