Problem
Agents can loop or call expensive models after budget authority is gone; the invoice arrives after the damage.
Most cost tools show the invoice. This one can refuse the call when the budget is gone.
Project information
- Category: Agent cost control · FinOps · Reliability
- Focus: Pre-call budgets, fail-closed deny, spend attribution
- For: Teams with shared agent budgets and surprise invoices
- Role: Reference system behind reliability engagements
Overview
Dashboards explain spend after the damage. Agent FinOps answers a harder question: can the system refuse an expensive call when the budget is already gone?
The outcome teams care about is simple. Hierarchical limits. Clear allow or deny. Attribution you can act on during the run, not only at month-end.
Work with me:
Mechanism
Policy checks budget, actor, workflow, and model request before spend happens, then fails closed when enforcement is unavailable.
What It Proves
Cost control can be active runtime enforcement, not a passive dashboard after the bill is generated.
Engagement Relevance
Useful for diagnostics and sprints where agent workflows need explicit spend authority and operator-visible limits.
CTA
Use this pattern to scope a diagnostic or one fail-closed systems sprint.
Send a problem briefTechnical evidence
Budget reservation denied before the provider call (reference system example).
HTTP 402 Payment Required
available_budget_usd: 0.20
projected_spend_usd: 0.50
allowed: false
avoided_cost_usd: 0.50
provider_call: not executed
event: budget_reservation_deniedKnown limitation: Redis-backed enforcement is required for production fail-closed behavior; in-memory mode is for local development only.
Goal
Make agent spend controllable before it hurts. Hard stops. Readable attribution. An operator view that treats cost as a reliability concern, not a finance afterthought.
What it does
- Pre-call enforcement: budget checks run before expensive model or tool work proceeds.
- Fail-closed posture: if enforcement cannot be trusted, the safe default is deny, not hope.
- Hierarchical limits: org and team envelopes so spend has an owner, not only a chart.
- Spend attribution: which feature, actor, or agent drove the cost.
How operators use it
Built for teams that already feel loop risk, fan-out, or surprise invoices. Cost decisions stay visible while the agent works, not buried in a monthly report.
- Clear outcomes: allow vs deny is part of the operator story, not a silent backend detail.
- Pause and resume: runs stop when limits trip; they continue after human or policy review.
- Entry wedge: a practical first step into broader reliability work (gates, memory, evidence).
Where it fits
- Use cases: multi-agent products, LLM gateways, internal copilots with shared budgets.
- Signal: especially useful when dashboards alone have not stopped overspend.
What I built
I built Agent FinOps as cost control on the execution path: enforcement timing, operator clarity, and a clean link into the rest of a reliability practice. Not another reporting dashboard.
Why it matters for clients
AI systems get safer when cost control sits in the execution path, not only in analytics after the bill arrives. That is the pattern audits and pilots install in your stack.