A single governed agent sits at the terminal node of ACME's showcase workflow — synthesizing every upstream hand-off into an auditable sign-off, and escalating to frontier reasoning when the stakes demand it.
When 367 agents act on your behalf, the question is never "did they work?" — it is "can you prove what they did, and who is accountable for the result?" The compliance-officer agent exists to answer that question by design, not after the fact.
The compliance officer is one identity inside a real, live zero-trust deployment — mapped like-for-like onto ACME's actual org and enforced on every call.
The directory behind the fleet is not synthetic: 120,120 humans and 205,034 non-human identities across 28 departments, 10 regions and 626 roles, synced from a real Samba-AD tree into the control-plane identity store.
Every governed call the compliance officer makes crosses the enforcement plane before it touches data. Deny-by-default is the resting state; an allow is an explicit, signed event.
Validates the tenant X-Tiresias-Api-Key and the per-call X-Capability-Token, evaluates policy, and only then forwards the request verbatim to the agent's own soul-svc tenant backend.
Each agent is its own soul-svc tenant with FORCE row-level security on Cloud SQL. The compliance officer cannot read another agent's memory — even a postgres superuser is subject to forced RLS.
The compliance officer's calls are observable on the same published surfaces an auditor would use to check them — right now, on the acme-demo host.
| Surface | Where | What it proves |
|---|---|---|
| PEP · governed data plane + /admin | :8343 | every allow/deny + audit chain |
| Asphodel management console | :3001 | agents, memory / CoT review, policy |
| Grafana · ROI economics board | :3000 | per-model $/1M-token cost |
| Loki · SoulWatch SIEM sink | :3100 | behavioral drift + syslog audit |
| Promtail · syslog ingest | :1514 | RFC5424 SIEM feed |
The showcase workflow is a 14-edge topological hand-off DAG. Upstream agents research, analyze and draft; the compliance officer is where their outputs converge for sign-off.
Each showcase agent runs with a home tier, a ceiling and a floor. The compliance officer's keyword-escalation triggers on the highest-stakes signals — +incident, +breach, +critical — and routes the reasoning up to the frontier model.
On its schedule, the compliance officer runs a regulatory-watch brief. The job is not special-cased — it travels the exact governed path every agent uses, and lands as a durable, isolated write.
This agent's authority rests on the same machinery that watches it. Its sign-offs are only as trustworthy as the chain that records them — and that chain is cryptographic, durable, and continuously monitored.
Every governed call is signed into a SoulKey Ed25519 audit chain with a durable WAL at /data/audit-wal, then shipped to the control plane. Verify it from the PEP /admin console.
Behavioral drift trips SoulWatch, which auto-quarantines the offending agent via the shared postgres store. A compliance officer that started misbehaving would be contained, not trusted.
Part of this agent's job is ROI synthesis — and the Grafana ROI board (uid acme-roi-economics) gives it real per-call economics to reconcile. The frontier tier it escalates to is the most expensive line by three orders of magnitude.
| Model tier | Provider | $/call | Relative |
|---|---|---|---|
| opus-4.8 (escalation ceiling) | OpenRouter paid | $0.19 | |
| sonnet-4.5 | OpenRouter paid | $0.03 | |
| deepseek-v4-pro | Ollama Cloud | $0.002 | |
| gpt-oss:120b | Ollama Cloud | $0.0015 | |
| Llama-3.3-70B | HuggingFace | $0.0012 | |
| gemma-4-31b:free | OpenRouter free | $0 |
The compliance officer is what a governed workforce looks like when it takes its own governance seriously: a terminal sign-off node that escalates to frontier reasoning under pressure, writes into a signed chain it cannot forge, and reconciles the economics of the whole workflow — every call deny-by-default, every action provable.