Config
Metrics
Risk map · likelihood × impact
Each dot is one agent sessionControls
Per control, since start| Control | Runs | Acted | p50 | p95 |
|---|
Gateway performance
From the read API, last 24 h| Path | Runs | p50 | p95 | p99 |
|---|
Sessions
Highest expected loss first| Session | P | C | Exp. loss | Level | Proposed action |
|---|
- Likelihood · P, 0 to 1
- How likely it is that the session has already gone wrong. Every session starts at 0.02. Each warning sign pushes it up: a blocked or redacted call, a request outside the task, a skipped check such as sanctions screening, a repeated write.
- Impact · C, 1 to 10
- How bad a step is if it was wrong: read 1, external call 4, write 5, create a client 8, run code 9, delete 10. The chart plots each session's worst step that actually ran.
- Expected loss · Σ P × C
- Adds up likelihood × impact over every step that ran. Blocked steps add nothing but raise the likelihood of later ones. It sets the level: medium from 3, high from 8 (a human must approve further actions), critical from 15 (the session is halted).
Session budget
80%
Audit log
Newest first · excerpts are sanitized| Time | Decision | Rule | Excerpt |
|---|
Tests
Suite by area
Each square is one test · hover for details| Area | What it proves | Passed | Cases |
|---|
Agent outcome checks
Recorded replays on synthetic dataOutcome check details
| Field | Approved | Actually saved | Check |
|---|