Should an agent guess or ask?
Ask. A question is cheaper than rework.
- Hypothesis
- An agent that stops to ask about unclear product intent produces less rework than one that fills the gap with a plausible guess.
- Method
- We gave the DayStorm Factory the same set of real issues from a large open-source frontend twice, weeks apart, in a clean room without access to earlier work. Between the two runs, the Factory gained a step that checks each spec against the codebase before the build and turns open product questions into explicit questions instead of assumptions.
- Finding
- In the first run, every change needed rework after the build. In the second run, every change was built and merged on the first attempt, with no human touching the code. One fix came out identical to the one the project's maintainers later made themselves.
An agent that guesses is fast until the guess is wrong. Then someone reviews a pull request that solves a slightly different problem, writes a correction, and waits for the next round.
The Factory takes the other route. Before a build starts, it checks the spec against the codebase. Where product intent is unclear, it asks: as a draft pull request with its question, where the answer is recorded too. You decide how often it may ask.