Skip to main content

Agents overview

dreeft currently ships three live agents, each tied to a stage of delivery. When a card enters one of these columns, the agent works with project context and writes results back to the card.

AgentRoleTypical outputs
AnalystClarifies requirements and scopePRD-style documentation, or forms when more input is needed
ArchitectDesigns the approach and breaks work downADR, stack notes, suggested tasks, task specs
CoderImplements in a sandbox and ships codeBranch, commits, Pull Request artifact

Planned QA stage

A dedicated QA page still exists in these docs because the product direction includes it, but QA is not a live default agent in the current board flow. Verification today happens through:

  • checks run by Coder
  • human review in GitHub
  • optional preview artifacts when preview infrastructure is enabled

How live agents are triggered

  1. You move a card into the agent’s column (or create it there).
  2. The engine checks configuration and handoff context.
  3. The agent runs and updates status from queued to processing.
  4. If more information is needed, the card can pause in wait_for_input.
  5. When the output is ready, the card usually lands in wait_for_review so a human can approve the next step.

Context and project knowledge

Agents are not isolated prompts. They work with project context such as:

  • prior artifacts on the card
  • project settings
  • repository information when Git is configured
  • codebase context gathered during the run

Review-driven handoffs

The current release is intentionally human-in-the-loop:

  • Analyst produces a PRD you review before sending the work forward.
  • Architect produces ADRs and suggested tasks you review before starting implementation.
  • Coder produces a PR you review in GitHub before merge.

Deep dives