Agent Cost Alerts That Actually Catch Runaway Spend

Monthly budget alerts are too slow

A monthly AI budget alert tells you what already happened. That works for stable SaaS usage. It does not work for agents.

Agent spend can move fast. A broken retry loop, a large context file loaded on every heartbeat, or a model fallback that silently routes to an expensive tier can turn a normal day into a bill spike before anyone sees the invoice.

The right alerts watch behavior, not just total spend.

Alert on spend rate

The first useful alert is spend per hour compared with baseline. If an agent usually costs $0.40/hour and suddenly costs $6/hour, you want to know now, not when the monthly threshold is crossed.

Rate alerts catch runaway loops, stuck background jobs, and sudden traffic spikes. They also avoid false comfort from low monthly spend early in a billing cycle.

Alert on retry volume

Retries are one of the easiest ways to waste LLM budget. A tool fails, the agent asks again, the same tool fails, and the loop repeats with a growing conversation history.

Track retry count by task and by tool. Alert when retries per task exceed a sane cap, especially when the model keeps receiving the same error. That usually means the agent is stuck, not thinking.

Alert on context growth

Context growth is quiet. The output looks normal, but every call gets more expensive as session history, memory files, tool definitions, and fetched documents accumulate.

A good alert watches input tokens per call. If the median input size doubles over a few hours, something changed. Maybe a prompt started loading a huge file. Maybe a tool began returning full records instead of summaries. Maybe a session is no longer being summarized.

Alert on model mix

Many teams configure fallbacks: use a cheaper model first, then route to a larger model when the call fails. That is sensible until a bad prompt or provider issue sends every call to the expensive fallback.

Watch the percentage of calls by model. If Opus or GPT-5.5 usage jumps from 5% to 60%, you want an alert tied to the routing change, not the eventual bill.

Alert on idle work

Heartbeat loops, scheduled scans, and background monitors should have predictable costs. They are also easy to forget because no human is actively chatting with them.

Set a separate budget for idle automation. If background work exceeds its normal daily envelope, investigate before it crowds out real usage.

The bottom line

Agent cost control is operational monitoring. The useful alerts tell you which behavior changed: rate, retries, context, model mix, or idle work.

Clawback shows those patterns at the agent-call level so you can catch runaway spend while it is still small.