Topic 3
Name the failures, then count them
Read first, categorise second. Write a plain note for each failure in your own words, then group the notes into failure modes and count each one. AI Engineering (Ch. 6, "Agent Failure Modes and Evaluation") puts the method in one line for agents: identify the failure modes, and measure how often each happens. The same works for any application.
Predict first
Twenty failing replies, read one by one
Illustrative notes from a review session
- Asked about invoices, got the generic billing FAQ: invoice page wasn't retrieved
- Said exports include attachments; docs say they don't
- SSO setup page never came back from search
- Question was about cancelling, reply explained pausing
- Quoted the old price, $8; docs say $10
- Offline mode question: retrieved the mobile install page instead
- Refused 'how do I delete my data?' as a privacy question
- Six paragraphs for a yes/no question
- No passage about team seats retrieved, so it guessed
- Invented a 'Pro Plus' plan
- Answered how to add a user when asked how to remove one
- Two-factor reset page missing from results
- 'Unfortunately, as I've said…' to a first-time question
- Said refunds take 3 days; docs say 5–7
- Searched 'API limits', got rate-limit page for the old API
- Refused to explain the data retention policy
- Replied about the desktop app; user was on web
- Wrong menu path: Settings → Billing doesn't exist on web
- Long list of every plan when asked for one price
- Webhook docs not retrieved; answer was generic
Twenty failing replies, read one by one, fall into six failure modes. The two biggest, retrieval that missed the right page and wrong facts, account for 60% of the failures. Failures often concentrate like this, so a short reading session can point at most of the problem.
The book's agent chapter shows what a taxonomy can look like for a more complex system: planning failures, such as an invalid tool or wrong parameters; tool failures, where the right tool gives a wrong output; and efficiency problems, too many steps or too much cost. Your own categories will be specific to your application, and that's the point.
Quick check