Use case / Tool-calling agents

See what the model asked the world to do.

Models pick tools, invent arguments and loop. Record every call with its payload, result and cost so a rejected schema or a runaway loop stands out immediately.

Start tracing free Read the docs

Arguments: Recorded
Loops: Detected
Cost: Per call

Instrument it

See the arguments the model chose.

Wrap each tool and the exact payload, the result and every retry are recorded, so loops and rejected schemas stand out.

$ npm install tracify-sdk

import { observe, traceAgent } from "tracify-sdk";

const issueRefund = observe(async function issueRefund(args: { amount: number; currency: string }) {
  return payments.refund(args);
}, { type: "tool" });

export const agent = traceAgent(async (request: string) => {
  const call = await pickTool(request);
  return issueRefund(call.args);
});

Payload inspector

What the model asked the world to do.

Start from what the user asked and which tool the model reached for.
Selection is its own span, so wrong tool and wrong arguments are separate problems.
Schema failures show the exact payload that was rejected.
Repeated repairs are the loop. Count them and see what each one cost.
The final result is checked against the call that produced it.