Skip to content

/

Engineering

Why we record tool calls, not just model calls

Half of the bugs we see in agents are in the tools around the model. Tracing only the model hides them.

T

Tomás Reyes

8 min

Chart illustration for why we record tool calls not just model calls

A model call has a clean contract: text in, text out. A tool call is a network request, a database query or a side effect, and it is where most production failures live.

What a tool span stores

  • The arguments the model chose

  • The raw response, redacted by your rules

  • Retries and timing, drawn to scale

The payoff

When a run goes wrong you can see whether the model picked the wrong tool or the right tool failed. Those need different fixes.

Create a free website with Framer, the website builder loved by startups, designers and agencies.