This is the queue as it arrives — before any agent has looked at it. A bank return code, an amount, and a memo somebody typed. Working out what each one means, one at a time, is the job.
Each case is triaged, investigated by specialists chosen for its type, diagnosed by an agent that holds no tools, then passed to a deterministic risk gate. Click any row for the reasoning. Inject a fault to watch the guardrails contain it.
★ marks a deliberately engineered case: a refusal, a prompt injection, and a reference to a record that does not exist.
Every agent runs under its own Google Cloud service account through impersonation. This page has each one attempt a real BigQuery read under its own identity — specialists reach their own table and no other, and the diagnosis agent is refused. That is IAM, not a prompt.
The published agent catalog, persisted in Firestore. Capability, version, identity and tool scope — what another team discovers when looking for an approved agent.
Escalated cases, highest value first, each with the reason it could not be resolved and what would have to change to clear it. A scheduled sweeper re-examines these and re-opens the ones whose blocker has cleared — sometimes weeks later.