← Back to the atlas
Testing & agents

What years as an SDET taught me about building agents

In progress Draft date

The habits that make failures reproducible and success measurable matter even more when software chooses its own next step

Working outline — full essay forthcoming

Rough outline

Define success before execution

  • Separate a completed action from a completed task
  • Specify acceptable outcomes, forbidden outcomes, and escalation conditions

Carry the debugging discipline forward

  • Capture inputs, tool responses, state transitions, and environment details
  • Use a small reproducible failure before reaching for a larger evaluation suite

Adapt the test strategy

  • Repeat runs to understand variability instead of treating one pass as evidence of reliability
  • Separate deterministic tool contracts from judgment-based outcome checks

Ground it in experience

  • Add a specific SDET incident and the diagnostic technique that resolved it
  • Show how that technique applies to an agent workflow, including where the analogy breaks