What information should a coding agent persist between steps?

Example Answer

The way I'd think about it is this: The agent should persist the task goal, relevant file and symbol references, important constraints, what it has already changed, what validation has run, and which assumptions remain unresolved. That is the durable state needed to continue safely.

I do not want it persisting every internal thought. Too much scratch reasoning becomes noise and can anchor the agent on bad early ideas. The persistent state should help execution and review, not just preserve verbosity.

Good persisted state makes reruns, human review, and safe resumption much easier.

What matters in an interview is not only knowing the definition, but being able to connect it back to how it changes modeling, evaluation, or deployment decisions in practice.

Common Poor Answer

A weak answer is saying the agent should save the full conversation so it never forgets anything. Durable workflow state is more useful than raw chat history.