Guide
Turn a failing AI workflow into a practical debugging checklist and a clear support ticket.
Start with one reproducible failure
A screenshot of a bad answer is rarely enough to identify the cause. Keep the input, the source documents and the tool response together. Write down the expected behavior in a sentence someone else can judge. Replay the smallest case that still fails before changing multiple settings.
Find the failing layer
Agent failures often start before the final model response. Missing documents belong to retrieval, expired tokens belong to the integration, and repeated write actions belong to orchestration. Test each layer separately with harmless sample data so a prompt change does not conceal a connection problem.
Build a regression set
Keep a short collection of real failures, normal requests and deliberately incomplete inputs. After each change, run the entire set and record what improved and what regressed. A fluent answer is not evidence that the agent followed your business rules.
How the result is made
Rule-based troubleshooting paths selected by failure category; all inputs are processed in your browser.