PullPlane's Research agent performs a read-only investigation of the repository and card spec, then returns concrete findings for the team or a later coding step.
A read-only research run
Research runs use Claude in ultracode mode and are instructed not to modify files. The run investigates relevant code paths, implementation options, risks, edge cases, and a verification strategy within a single research session.
The result is a written set of research findings rather than a branch of code changes. That separation makes the evidence and recommendation reviewable before implementation begins.
Questions the Research agent can answer
Use research when the team has not yet agreed on the right implementation boundary or when a change crosses unfamiliar components. A specific question and bounded scope produce a more useful handoff than an open-ended request to inspect everything.
- Which files and functions control the behavior?
- What approaches fit the existing architecture, and what are their trade-offs?
- Which failure modes and regression checks should shape the implementation?
Turn findings into an implementation brief
Review the recommendation, resolve any human decisions, and add the accepted approach to the card's acceptance criteria. A later Claude or Codex step can then implement against that shared context without pretending the research itself changed the product.