Install CAUM beside one recurring AI agent workflow. CAUM returns zero-semantic receipts and a live dashboard for loops, retries, workflow memory, review boundaries, token exposure, and cost exposure without reading prompts, files, messages, source code, or business payloads.
CAUM is narrow by design. It does not try to be a general tracing suite or a truth detector. It measures whether agent work still has healthy structural movement, where it first stops converting cleanly, and what budget exposure sits after that passive review boundary.
T1-T5 health tiers for each run, with review priority and profile-aware calibration.
The first structural point where a run should become reviewable, with tail tokens and cost after that point.
A zero-semantic readout of whether observed steps are still turning into structural movement.
Customer-scoped review cards for recurrent fingerprints, conversion drift, and long-horizon rhythm across repeated sessions.
A one-time receipt shows the structural loop or retry pattern. CAUM Live keeps that evidence attached to running agents, where repeated loops, retry storms, reasoning stalls, tool churn, and workflow-level memory can accumulate across days or teams.
CAUM Live can compare customer-marked before/after cohorts after a team applies its own retry ceiling, handoff rule, or exit contract. The output is an observed structural exposure delta, not a realized savings claim.
Three passive policies were tested as before/after structural cohorts. CAUM observed lower reviewable exposure after the policy marker was applied.
| Policy | Exposure delta | Reviewable delta | Token delta |
|---|---|---|---|
| Retry error ceilingMarks repeated failed retries before they become hidden churn. | 50.4% | $119.90 | 358,400 |
| Handoff bounce limitSurfaces repeated routing between actors or tools. | 79.0% | $173.30 | 435,000 |
| Reasoning-to-action exit contractDetects long reasoning runs that stop becoming action. | 78.4% | $289.63 | 1,051,800 |
Live events can include agent identity and delegation metadata when available. CAUM hashes identity fields in returned evidence and ignores sensitive content fields.
{
"session_id": "deploy-184",
"agent_id": "agent-42",
"identity_sub": "spiffe://acme/prod/agent/42",
"scopes": ["tools:read", "tools:execute"],
"event": {
"event": "tool_call",
"tool": "bash",
"status": "completed",
"input_tokens": 420,
"cost_usd": 0.013
}
}These plans buy continuous structural observability and onboarding for agent runs. CAUM Receipt remains the low-friction entry point; CAUM Live is for recurring agent spend and operational review.
For one builder who wants CAUM on coding-agent runs without an enterprise process.
For teams running agents often enough that repeated loops, retries, and reviewable cost exposure already matter.
For higher-volume pilots that need a more guided integration path and operator review.
CAUM observes structural evidence only. It does not judge whether an answer is true, does not make content-truth claims, and does not block agents.
Use CAUM Receipt to enter the workflow. Use CAUM Live when the same agent behavior repeats often enough that reviewable structural exposure becomes a recurring operating problem.