Operating Note 003: Spend Policies Per Agent Role
Operating Note 003 (3 Sep 2026): Spend governance for multi-agent stacks. Builds on Operating Note 002 (task classes + roles) and the cost/ROI series hub. Company OS decides what may run; this note is how much each role may spend doing it.
Thesis
Most agent orgs fail cost governance the same way they fail model routing: one API key, one budget, no role boundaries. Research loops eat frontier tokens. Implementers retry on expensive models because nobody capped the lane. Publishers ship without a separate approval path because spend and risk share the same wallet.
The fix is not a finance spreadsheet — it is spend policy per agent role: explicit ceilings, model tiers, and approval gates that match how work actually flows. Same shape as Company OS intent → policy → outcome, applied to euros and tokens instead of org charts.
Right agent, right model (Note 002) tells you which model fits the task. Spend policy tells you how much that role may consume before a human must intervene.
Practice
Define four roles with different spend posture and approval rules. Use your router labels (OpenRouter, LiteLLM, or static allowlist) — the policy is role-first, vendor-agnostic.
| Role | Owns | Model tier | Spend posture | Approval |
|---|---|---|---|---|
| Researcher | Triage, extract, outline, web gather | Free / tiny (class A) | High volume, low unit cost; hard cap on concurrent runs | None for read-only; block external writes |
| Implementer | Code, schema drafts, tool calls with tests | Cheap mid-tier (class B) | Moderate cap per workflow; retries budgeted | Auto for repo branches; human for main/prod |
| Critic | QA another agent’s output, risky diffs | Stronger / different model than producer (class C) | Lower volume, higher unit cost; cap per review batch | None if read-only critique; escalate on block |
| Publisher | External publish, email, deploy, IndexNow | Strong model (class D) | Lowest autonomous cap — smallest wallet | Human required before any external side effect |
Policy template (copy and adapt)
For each role, document in your runbook or Company OS config:
- Allowed models — smallest set that clears the eval bar for that lane
- Spend ceiling — per day, per workflow, or per accepted outcome (pick one unit and stick to it)
- Hard stop behavior — pause queue, alert human, never silently downgrade to a weaker model on D-class work
- Approval hook — who must click before publisher runs
Start with three ceilings, not twelve: researcher volume cap, implementer per-PR cap, publisher near-zero autonomous spend.
Wiring in practice
- Tag every agent invocation with role in logs and router metadata
- Split API keys or sub-accounts per role if your gateway supports it — otherwise enforce in orchestration
- Weekly review: spend by role vs accepted outcomes (model watch feeds promote/demote; spend policy feeds cap adjustments)
- Never let publisher share implementer’s budget — that is how draft posts go live without review
The €35/mo content org already separates infra from model spend and gates external publish. This note generalizes that pattern to any multi-agent stack.
Proof
Evidence from this studio stack (no new invented figures):
- Infra band ~€35/mo for the autonomous content org — model/API spend tracked separately so routing changes do not hide in hosting bills (case study)
- Producer vs critic models in the agent stack — different models for generation and review, not one frontier call for both (stack writeup)
- Class D + human approval already documented in Note 002 for irreversible external actions — spend policy makes that economic, not just procedural
- OpenRouter routing tables with promote/demote — spend by lane is measurable when roles map to routes (Cursor example)
If you cannot attribute spend to a role this week, your policy is not wired yet — fix logging before lowering caps.
Next
Shipped in this lane:
- Operating Note 002 — task classes + role vs model
- This note — spend ceilings per role
Still planned in the series hub: failure-priced ROI ledger week, domain-specific eval fixtures, autonomous content org “done” definition.
Series hub: Agent cost & model selection — start here.
Questions on agent spend limits or multi-agent governance — contact me with a short note on your stack.