Build reliable AI systems, one boundary at a time
Start with a task and its failure cases. These guides use explicit examples and distinguish reference designs from measured results. Read in this order, or choose the problem you are working on.
1. Identify what can go wrong
Turn an ambiguous task into explicit failure cases.
Five agent failure modes to test before release →3 min read · Practical guide
2. Check outcomes independently
Build a rubric, validate scores, and review disagreements.
Build an agent evaluation suite: cases, rubrics, and release decisions →4 min read · Practical guide
3. Bound the execution
Separate language interpretation from authorized actions and recovery.
Designing a bounded agent workflow: states, tools, and recovery →3 min read · Reference design
4. Inspect a voice interaction
Measure meaningful audio response and test stream completion.
Where voice-agent latency goes: a turn-by-turn budget →3 min read · Practical guide
5. Review an AI-assisted change
Work through the sentence-chunking defect and its regression cases.
One bug with a coding agent: the contract, the mistake, and the review →3 min read · Worked example