An agent is not behaving as expected

Most agent problems are context problems, not instruction problems. Work in this order.

1. Read the trace

The trace shows what the agent read, which tools it called, and why it stopped. Read it before changing anything. An agent that never read the relevant record has a data problem, not a prompt problem.

2. Check what it could see

Confirm the source holding the answer is connected, synced recently, and in scope for this agent. An agent cannot use what it was never given.

3. Check the tools

If the agent should have acted and did not, confirm the tool is attached and that the run did not stop at a boundary you set.

4. Only then change the brief

Change one thing at a time and rerun against the same case, so you can tell which change did what.

If the behaviour is inconsistent between identical runs, capture two traces and contact support with both.