Free diagnostic

Find the layer that is actually failing.

Answer eight evidence-based checks in order. The first failed check usually identifies the layer to investigate before changing prompts, models, or production configuration.

~10 minutesNo signupNothing is uploaded
01Can the deployed runtime reach the endpoint?

Test DNS, TLS, route, and response status from the same environment where the agent runs—not only from a laptop.

02Does the client discover the expected tool and schema?

Confirm the exact tool name, required inputs, types, and description exposed to the model at runtime.

03Does an authorized raw request succeed without the agent?

Separate token issuance, audience, scope, redirect, and permission problems from agent-selection behavior.

04Is the deployed identity and configuration the intended one?

Verify secret presence, environment names, callback URLs, runtime versions, regions, and service-account permissions.

05Does direct tool execution return the expected result?

Invoke the tool with a minimal valid payload and inspect status, latency, output shape, side effects, and error propagation.

06Does the model select the tool when the task requires it?

Inspect the actual prompt, available-tool list, tool-choice settings, and competing instructions in the failing run.

07Can the agent consume the tool result?

Check JSON validity, size limits, required fields, error envelopes, content types, and whether the result is attached to the right turn.

08Does the full workflow prove the intended outcome?

Run one production-like path through retries, state transitions, callbacks, persistence, and the final user-visible result.