Model your AI agent step by step — the model each step calls, its token profile, its loops and retries, and the external tools it depends on. AgentObserve maps the dependency graph, projects per-run and monthly spend, and scores the guardrails your CFO will ask about. Nothing is uploaded; it all runs in your browser.
How often the whole agent runs, and how much of the repeated input (system prompt, tool schemas, RAG context) is served from prompt cache.
One row per LLM call in your agent. Calls/run captures loops and retries — a 10-iteration reasoning loop is 10 calls; a tool-use step that averages 3 round-trips is 3. Set Tool / API to the external dependency a step invokes (search, DB, a paid API) and its per-call cost if any.
| Step | Model | In tok | Out tok | Calls/run | Tool / API | Tool $/call | $ / run | $ / month |
|---|
Each step sized by its share of per-run cost, with the model it calls and the external tools it depends on. The widest bars are where your spend — and your failure risk — concentrate.
The controls that keep agent spend from running away. Public LLM providers won't cap your usage — these are the guardrails enterprise FinOps teams put in place.