Agent Cost Monitoring: What a Cost Spike Usually Means

A 200-task benchmark published in April 2026 found that 90.8% of an AI agent’s retries were wasted: 466 of 513 retry attempts hit a tool name the model had hallucinated, a permanent error no amount of retrying could fix (Towards Data Science).

Every one of those retries is a billed API call. A team watching only total spend sees those wasted retries as a cost line moving the wrong direction, not as a hallucination problem with a name.

This is the pattern behind most agent cost monitoring gaps: the dashboard shows the number going up, not why. Cost is often the first signal a production agent gives when something breaks, and it usually arrives before a user complains or a support ticket opens.

What Rising Agent API Costs Usually Mean

Three failure modes inside the agent produce a cost signature before they produce a visible quality problem.

The first is retry storms on non-retryable errors, the pattern in the benchmark above. A hallucinated tool name or a malformed argument fails every time, and a retry policy built for transient errors like timeouts burns budget on a call that was never going to succeed.

The second is context accumulation. Augment Code’s analysis of agent loop token costs found that each step’s tool output and reasoning trace gets carried into every later call, so a 20-step loop generates roughly 10 times the cumulative input tokens a flat per-step estimate would suggest.

In their worked example, a naive 10-step file-reading agent cost about 43 times more than a single-pass version of the same task.

The third is a runaway tool loop: the agent calls the same tool repeatedly with similar arguments, reasoning over a result that never resolves the task. Each iteration bills tokens whether or not it moves the agent closer to done.

There is a fourth that never comes through your repo. The model provider updates the model underneath you, responses get more verbose, and per-session cost climbs with nothing in your diff to explain it.

None of these show up in an error rate. The agent completes the task, or times out cleanly, and the cost is the only signal that something is wrong.

Telling Benign Growth From a Reliability Problem

More users spending more money looks similar to a misbehaving agent spending more money, at least on a total-spend graph. The difference shows up at the session level.

Benign growth moves with traffic: session count and total cost rise together, and average cost per session stays flat. A reliability problem moves independent of traffic: cost per session climbs on its own, or a small number of sessions account for a disproportionate share of new spend.

Each failure mode leaves a trace-level fingerprint that doesn’t require knowing the agent’s error rate ahead of time. A tool invoked repeatedly with near-identical arguments that keep failing echoes the retry-storm pattern in the benchmark above; the same shape where every call succeeds but the task gets no closer to done is a runaway loop. A prompt size that climbs step over step within a session echoes the context-accumulation pattern Augment Code measured. Output tokens jumping across every session at once, with no deploy on your side, points at a provider update, not your code.

Setting Up Agent Cost Monitoring That Catches Quality Problems

A cost anomaly is a useful alert only if it points at a cause. Session-level cost tracked alongside the signals already watching for hallucination and tool failure turns a spend spike into a specific, actionable finding instead of a line item someone has to go investigate by hand.

The New Stack’s reporting on agentic system failures put it plainly: “Systems that cannot answer questions about which actions they took, why those actions were taken, how much each action cost and how/why each action failed have not moved past the demo stage, regardless of the quality of the output.”

Correlating cost with quality is what closes that gap. Tessary reads the trend in cost per session alongside the trend across grader verdicts, rather than any single spike; a grader misfires at a stable rate, so cost and a quality signal moving together marks a real change. When they do, the same graders that already flag tool errors and hallucinations show which one fired on the expensive sessions, and the correlated alert routes to Slack, PagerDuty, or Linear instead of a dashboard nobody is watching.

The correlation still isn’t the cause. The last step is tracing the failure back to the change that started it: the prompt edit that kicked off the retry storm, the tool update that broke a schema and turned every call into three, the provider update that raised token counts with no deploy on your side. When your agent starts costing more in production, Tessary tells you what caused it, in time to fix it before the bill compounds.

Tessary runs this over your existing traces, whether they arrive through OTLP or get pulled from a connected Langfuse or Braintrust source. Your Langfuse dashboard will show the cost line and the score line both moving; it won’t say which change made them move. Tessary reads the same traces and answers that question. Setting up agent monitoring and alert routing covers the wiring. Connecting a source takes a few minutes, with no new instrumentation and no credit card.

Find out what caused your last cost spike