research

To conduct objective codebase research without confirmation bias, use parallel specialist subagents that research isolated questions while withholding project goals. This prevents product intent from leaking into findings, producing factual, reviewable research summaries.

You synthesize parallel research findings by dispatching subagents per numbered question and consolidating their outputs. The process aggregates per-question reports into a unified research summary file, running structured review rounds to ensure quality before human approval.

The best way to enforce research isolation in agent orchestration is to gate subagent prompts, ensuring they receive only approved questions without project goals files. This strict gating prevents context leakage and keeps findings tied strictly to specified research questions.

Yes, you can run web searches and codebase reads concurrently for software research. Multi-type research supports parallel execution of local code analysis with file references and external web searches with URLs, enabling hybrid comparisons between local and external sources.

Handle rejected research summaries by running targeted re-runs of individual researchers or synthesis rounds. The system preserves human feedback in dedicated review files, allowing you to dispatch specific subagents to correct isolated findings without redoing the entire research process.

Yes, you need an approved questions file to start automated research. The pipeline requires a questions file with an explicit approved status before dispatching any specialist subagents, ensuring research only begins on explicitly validated and numbered queries.