Probing depth and follow-ups

probing_depth on the agent and max_follow_ups_per_question on boundaries together control how hard the agent pushes within a node. Get this wrong and you either leave material on the table - light when the interview wanted depth - or burn out the respondent - deep on a six-question interview meant to take five minutes. This chapter is how to decide.

The three depths in practice

What each depth actually does at runtime: light issues at most one follow-up, only when the answer is visibly thin; moderate probes when it detects a surface answer or genuine curiosity; deep follows up several times until it hits the max_follow_ups_per_question ceiling or the completion_hint is satisfied.

Depth matters more than technique selection. A deep interview with a dull technique list outperforms a light interview with a perfect one.

max_follow_ups_per_question

The hard ceiling. Set it on the agent's boundaries field. The agent will not exceed this regardless of probing_depth.

Recommended starting points:

• light → 1
• moderate → 2–3
• deep → 4–5

Lower it for chat channels - respondents tire faster than in voice. The voice/chat differential roughly halves what's tolerable.

This is the integrator's safety valve. If a real run shows respondents dropping out mid-interview, lower the ceiling before you change anything else.

Writing completion_hints

Per-node probing.completion_hint field on add_blueprint_node. It's the agent's definition of "done" for this node - the agent will keep probing until the hint is satisfied or the follow-up ceiling is hit, whichever comes first. Without a hint, the agent falls back to its default sense of completeness, which is usually too lenient.

Three before/after rewrites:

Vague: "Get a good answer." Better: "The respondent has named at least one specific moment when the issue occurred and described what they did in response."

Vague: "Cover the topic of pricing." Better: "The respondent has stated whether the price felt fair, and either explained why or named what would have felt fair."

Vague: "Make sure they're satisfied." Better: "The respondent has rated their satisfaction and given at least one reason for the rating."

A tight hint reads like a quality bar, not a question. It names what the agent should have heard, not what to ask. Pair with the techniques list to shape how the agent gets there.

The depth/burden tradeoff

A trade you have to make consciously. Deeper probing yields richer transcripts but lowers completion rates and lengthens runs. Shallower probing is kinder to respondents but produces survey-shaped data, not conversation-shaped data.

Rule of thumb: pick the depth your real respondents will actually finish, not the one you wish they would. Run five interviews at your chosen depth and look at completion rates before scaling.

Use case rule of thumb

These are starting points. Adjust after five real interviews.

What to read next