Skip to content

How It Works

Summon isn’t just a set of agents — it’s a methodology. Every agent knows where it fits in the workflow, when to hand off to the next, and when to push back.

You → Cam (Discovery) → Archie (Architecture) → Tara+Sato (Build)
Vik+Tara+Pierrot (Review)
Grace (Sprint Boundary)

Lead: Cam — You describe what you want. Cam doesn’t let you get away with vague input. She probes, clarifies, and pressure-tests until the vision is concrete. Only then does the team move on.

Lead: Archie — Before anyone writes code, Archie designs the architecture and writes an ADR (Architecture Decision Record). Wei plays devil’s advocate, challenging assumptions. No implementation starts without an approved ADR.

Lead: Pat + Grace — Pat writes acceptance criteria and prioritizes the backlog. Grace organizes work into sprint waves, sized for Claude Code’s context window.

Lead: Tara → Sato — Strict TDD. Tara writes failing tests first. Sato makes them pass. One commit per issue, conventional commit format. No shortcuts.

Lead: Vik + Tara + Pierrot — Three parallel lenses, every time:

  • Vik checks simplicity and maintainability
  • Tara checks test quality and coverage
  • Pierrot checks security and compliance

Critical findings block the merge. This isn’t optional.

Every work item passes a 15-item checklist before closing:

  • Tests pass and cover the change
  • No security findings
  • Docs updated
  • ADR written (if architectural)
  • Board status updated

Lead: Grace — Retrospective, velocity tracking, and kaizen. What worked? What didn’t? What to improve? Then the next sprint starts with better process.

Claude Code has finite context. Summon accounts for this:

  • /handoff saves session state so you can resume later
  • /resume picks up exactly where you left off
  • Work is organized into waves — sized chunks that fit in one session
  • Commit frequently — uncommitted work is expensive to reconstruct