Writing ·
Agent cost runbook: when agent spend outpaces output in 2026
A cost runbook classifies the loop, contains the damage, and measures accepted output so routing and retry controls get tuned for the right failure mode.
By Youssef Hemimy · cost controls · agent reliability · runbooks
An agent cost runbook should tell you what to do when agent spend outpaces output. Start by classifying the loop, then contain the bad path, then tune the guardrails that control routing and retries. The goal is not to make every call cheaper; it is to make the harness spend money on useful work.
Detect
spend, retry, or token slope
Classify
loop, tool, model, or queue
Contain
route, pause, or degrade
Reconcile
accepted output vs. waste
Tune
budget, eval, and retry guard
Unit of review
Review cost by accepted output, not by raw model calls. Raw spend tells you the bill moved; accepted output tells you whether the work moved.
Classify before you change anything
Cost spikes usually come from one of four places: too many retries, an expensive routing choice, a stuck tool, or a queue that has become a backlog. If you change the model before classifying the loop, you may only hide the real problem.
| Signal | Likely action |
|---|---|
| Retry count rising | Shorten retry budget or fix the failing tool |
| Model mix drifting | Tighten routing rules |
| Queue depth climbing | Reduce intake or add capacity |
| Accepted output falling | Pause and inspect the workflow |
Contain the waste
Containment is about buying time. You can route to a cheaper lane, pause the workflow, or narrow the tool surface while you investigate. The right move depends on whether the loop is unhealthy, merely expensive, or actually dangerous.
Tune the guardrails
Once the incident is contained, tune the controls that failed to hold: retry budgets, routing thresholds, and the alert boundary that tells you the loop is drifting. The tuning goal is simple: keep accepted output steady while the waste falls.
This is the same reliability loop used in multi-model routing cost controls and the broader AgentOps Hardening model behind AgentOps.
FAQ
What is the first thing to check when cost rises?
Classify the loop first. You want to know whether the waste is coming from retries, routing, a stuck tool, or a queue that is backing up.
Why measure accepted output instead of raw spend?
Raw spend tells you the bill moved. Accepted output tells you whether the work moved. The runbook should optimize the second, not the first by itself.
Should every cost spike trigger a stop?
No. Some spikes need routing changes, some need shorter retry budgets, and some need a human pause. Stop only when the loop is clearly unhealthy or risky.
Where does this belong in AgentOps?
It is part of the operating telemetry layer. Cost is not just a finance metric; it is a reliability signal about whether the harness is doing too much work for too little output.
Building something that has to hold up?
We do this work for teams — agent reliability hardening, custom MCP servers, and full-stack AI systems built to survive production.