Skip to content
Current page: Research Before Code AI Agent Workflow

Put research before implementation

Use a read-only Research agent to map code paths, risks, and options before handing an agreed recommendation to a coding agent.

Separating investigation from coding gives the team a visible point to challenge assumptions and choose an approach before an agent changes the repository.

Frame a decision, not a scavenger hunt

State the outcome, scope, known constraints, open technical decision, and evidence the Research agent should gather. Repository-wide questions without a boundary make the handoff less actionable.

Read-only investigation

The Research agent inspects the repository and card context without modifying files. It returns relevant code paths, architectural context, options and trade-offs, failure modes, and a recommended implementation and verification approach.

Human checkpoint before build

A teammate reviews the findings, resolves product and architecture decisions, and updates the card. Only then should Claude or Codex implement the accepted approach. The research-build-review starter recipe captures that sequence as a reusable chain.

Bring every coding agent onto one shared plane.

Plan with your team, run Claude and Codex in isolated sandboxes, and review the work before it ships.

Start a project