← sources

We route our coding agents' questions by type of decision, not by risk

source: reddit (reddit:AI_Agents)published 2026-10-29 12:35ingested 2026-10-30 02:36

retained text

We're building a business system mostly with coding agents. When a task runs past what its spec covers, the agent stops and asks, with options, instead of guessing.

Then we route the question by what kind of decision it is:

* what the product should do, direction, an exception to a rule, or loosening a check: a person decides
* routine design calls inside rules we've already written down: answered without the owner, from those rules and earlier rulings

Every answer is recorded with who gave it, so later questions can point to earlier decisions instead of being asked again.

For scale: in the week to 26 October 2026, agents raised 762 design questions. Of the 737 answered, the owner answered 59.

The part I think matters most: "not sure which route this is" has to count as a reason to ask. Otherwise the agent ends up judging how risky the question is, which is exactly the judgment you were trying to keep with a person.

How are others drawing this line? By kind of decision, by cost of a mistake, something else? And if it's cost, what stops the agent from underestimating it?

document fetched 2026-10-30 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xiiris/we_route_our_coding_agents_questions_by_type_of