AI

Claude Cowork

Multi-agent cooperative work.

A pattern for getting several Claude instances to plan, build, review, and ship in parallel — with humans steering the meta. Where solo-agent work hits a ceiling, cowork doubles throughput.

  • Planner ↔ builder ↔ reviewer ↔ deployer loops
  • Shared context via tasks, plans, and project memory
  • Used inside AIBO, APPAI, and every Mat Siems engagement
Why

Solo agents hit a ceiling fast — context fills, focus drifts. Cowork breaks the wall: several Claude instances plan, build, and review in parallel, with you steering the meta instead of typing prompts.

How
  • One planner, one builder, one reviewer — strict role separation
  • Shared tasks + plans as the coordination substrate
  • Human signs off at the checkpoints that matter
Proof
Throughput vs. solo
2.4×
Cowork-built platforms
AIBO · APPAI · AIOS
Concurrent agents
up to 13
Cowork — four roles, one loop
Planner ↔ Builder ↔ Reviewer ↔ Deployer
Hover or tap a node to see details.
FAQ
Doesn't running four Claude instances cost four times as much?
Per-token, yes — but cowork gets the work done 2.4× faster in practice, and the reviewer/deployer catch issues that would cost more to fix downstream. Net: cheaper per shipped feature.
Do the four roles need different model tiers?
Often yes — Opus for the planner (hardest reasoning), Sonnet for builder/reviewer, Haiku for deployer status checks. Cowork is when you put the model budget where it matters most.
What's the human's role in cowork?
Steer the meta — confirm the plan, sign off at checkpoints, intervene if a role starts drifting. Not type prompts; not write code; not chase failures. Conductor, not session lead.
In production
  • AIOS V3 shipped in 5 days

    Planner ↔ builder ↔ reviewer ↔ deployer pattern moved AIOS from spec to V3 platform in five working days.

    See it
  • APPAI cockpit

    13-site orchestration cockpit built using cowork — concurrent agents on registry, build pipeline, and UI.

    See it
  • Throughput vs. solo

    2.4× measured throughput on non-trivial features when cowork is wired vs. a single Claude Code session.