Every call an agent makes to a model is recorded: what went in, what came back, which tools were invoked, how many tokens it burned, how long it took, and whether it failed. This is where you find out what your agents are actually doing.
When you'd use it #
A bad answer
A customer got something wrong. Open the exact inference and read what the model was given and what it returned.
Unexpected cost
Token usage per request shows which agent — and which prompt — is expensive.
Slow replies
Latency per request tells you whether to change model or trim the memory window.
What a record holds #
| Field | What it does |
|---|---|
| Agent and model | Which agent ran, and which model it called. |
| Status | Succeeded or failed. Failures carry the provider's error, which is usually an expired key, a rate limit or a context-length overflow. |
| Tokens | Input and output. The honest measure of what a conversation costs. |
| Latency | End-to-end time. What the customer experienced as waiting. |
| Prompt and response | The full exchange, including tool calls and their results. |
Where to go next #
AI agents overview
What an agent is, when it replies, and how a human takes the thread back.
Playground
Test an agent against real prompts before a customer ever reaches it.
Usage & limits
How much of each plan limit you have consumed this cycle.
Providers & models
The model providers and API keys your agents run inference against.