Improved token efficiency and focus in Deep Search

Deep Search can now use a dedicated subagent specialized for finding files relevant to your query in case a specific code domain needs to be researched. This subagent performs scoped searches and presents a summary to the main agent. Since the main Deep Search agent only needs to read this summary, this saves tokens in its context window.

This means Deep Search can explore more of your codebase while using fewer tokens, so you get more focused context and longer conversation. We expect this feature will significantly alleviate issues with deep single-question research - for multi-question, we just shipped auto-compaction.

We've already tried shipping this earlier this year and had to revert the change due to unforeseen issues with latency and agent behavior. This time, we've ironed out the wrinkles, made sure to scope the agent more narrowly and cap its maximal runtime with guaranteed results.