How to use MAST on your own system

The article gives a concrete recipe, which is the practical heart of the piece:

1. Collect execution traces. Instrument your system to save complete conversation histories, including all messages between agents and with the environment, across a variety of task types and difficulty levels.
2. Set up the MAST annotator. Use the published MAST repository as a base, and prompt a capable model with the full list of failure-mode definitions plus a few example traces (this is few-shot prompting).
3. Annotate your traces. Run them through the annotator to label each with the failure modes present, and manually validate a subset to check reliability.
4. Analyze the distribution. Ask which failures are most common, whether they cluster in one phase (coordination versus verification), and which agents propagate the most errors.
5. Design targeted interventions. Focus on the high-frequency failure types: add clarification strategies if agents proceed without asking; add domain-aware test cases if verification is weak; reorder agent responsibilities if specification issues dominate.
6. Re-run and compare. Use the same pipeline after your changes, and compare both the overall success rate and the shift in the failure distribution.

This turns failure analysis from guesswork into a repeatable engineering loop, the same way traditional software is profiled and optimized.