At fleet scale, the risk is not a runaway model — it is cost sprawl no one can account for. This walkthrough makes every governed call legible: routed, metered, and priced.
A general-purpose fleet drifts to the most expensive model that answers first. Multiply an unmanaged per-call cost across hundreds of agents running every day, and the spend is real long before anyone can attribute it.
No per-agent metering. No per-model attribution. Premium models used by default. The invoice arrives as one opaque line item — impossible to challenge, impossible to forecast.
Every call attributed to an agent, a department, and a model — with the dollar cost recorded at the point of use, not reconstructed after the fact.
367 governed agents run under zero-trust policy on a live GCP instance — not a sandbox. 360 fleet agents are mapped like-for-like to Acme's actual organization, plus a 7-agent showcase hand-off DAG.
Source: registry (367 agent tenants, postgres) · agent-manifest.json (360 fleet) · Samba-AD directory — 120,120 humans + 205,034 NHI, 28 departments, 10 regions, 626 roles
Every governed call passes the Policy Enforcement Point. On allow, the runner reports usage — tokens and dollars — which lands in a durable postgres table and surfaces on the ROI board. Nothing is estimated later.
Each row carries tenant, agent, model, tokens_in / tokens_out, and cost_usd, at hour and day granularity. That is the audit trail behind the invoice.
Source: tzt_usage schema · Postgres datasource acme-tzt-usage → Grafana :3000
The fleet runs as independent rolling batches — not a synchronized burst. The cadence is deliberate and measurable, which is exactly what makes the spend forecastable.
Each job is a full governed unit of work: domain brief → tiered LLM call → governed chain-of-thought + memory write → usage report.
Source: fleet_runner.py — FLEET_BATCH=6, INTERVAL=300s, MIN_GAP=10h (acme-fleet-runner, live)
Routing policy favors free and Ollama tiers by default. Premium paid models are reserved for genuine escalation — an incident, a breach, a critical keyword — not for routine work.
Source: config.json (9 tiers, 3 providers) · fleet_runner.py routing · personas.json escalation keywords
This is the ROI board's per-call economics. The spread between a premium model and a self-hosted one is not a rounding difference — it is more than two orders of magnitude.
| Model | Tier | Observed cost / call | USD |
|---|---|---|---|
| opus-4.8 | paid | $0.19 | |
| sonnet-4.5 | 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 | free | $0 |
Source: Grafana ROI dashboard (uid acme-roi-economics) · per-model $/1M-token economics from tzt_usage.cost_usd
Routing the fleet to the escalation ceiling instead of its free home would multiply the per-call cost by more than 100x — for identical work. Governance is what keeps the default cheap.
$0 – $0.0015
free / self-hosted tiers — where 1–2 jobs per agent per day actually run.
$0.19
opus-4.8 — justified only on a flagged incident, never as a default.
Source: config.json tiers · ROI board (uid acme-roi-economics)
The board surfaced a real saturation signal: hermes-405b:free returns chronic 429 rate-limit errors under fleet load. The free tier cannot carry production volume — which is precisely what makes the self-host break-even calculable.
Source: ROI dashboard free-tier saturation panel (uid acme-roi-economics)
Cost legibility rests on enforcement. No governed call reaches a model without a validated capability token, a deny-by-default policy check, and a signed entry in the audit chain. The usage row and the audit row describe the same event.
Source: PEP enforcement plane (acme-tiresias-zt, :8343) · signed audit chain → CP · tzt_usage
Attribution is not a monthly reconciliation exercise. It is recorded at the point of enforcement, for every one of the 367 agents, in the same durable store that powers the ROI board.
Source: Grafana ROI dashboard (uid acme-roi-economics) · tzt_usage.cost_usd · config.json tiers
367 agents, free-first routing, deny-by-default enforcement, and a per-call dollar figure recorded at the point of use. When finance asks what the AI workforce cost, the answer is a query — not an argument.