Reliability is mostly about knowing when not to answer. Build explicit escalation rules for an agent that takes consequential actions, and check that its confidence signal means anything.
What to build
- Build an agent that takes an action with real consequences — a refund, a config change.
- Define escalation triggers before you write the prompt: value thresholds, ambiguity, conflicting sources, low retrieval coverage.
- Implement the triggers as deterministic code, not as prompt instructions.
- Make the escalation payload complete: what was asked, what was found, what is uncertain, what it would have done.
- Test with an ambiguous request and confirm escalation rather than a guess.
- Sample fifty runs and check whether stated confidence tracks actual accuracy.
Done when
- Escalation triggers are enforced in code and cannot be argued away by the model.
- A human receiving an escalation has everything needed to decide without re-investigating.
- An ambiguous input escalates instead of resolving to the most likely reading.
- You have data on whether the confidence signal is calibrated.
The trap this exercise teaches
Escalating with only "I am not sure". The handoff must carry the reasoning and the evidence, or the human simply restarts the work.
Before you start
The theory behind this build is covered in Provenance, conflict and escalation. If any step below is unfamiliar, read that first — the exercise assumes it. Primary source: Anthropic — Building effective agents.
Check yourself against the exam
This exercise sits in Domain 5, which is 15% of the CCAR‑F exam. Once you have built it, run a domain drill in the study app and see whether the questions read differently.