The Coworkers Platform

A role-based multi-agent system: Claude instances that work as peers across the twelve knowledge-work domains — bio-research, customer-support, data, design, engineering, finance, human-resources, legal, marketing, operations, product-management, sales — messaging each other directly, gated by human approval where it matters, made visible through a physical desktop device.

The coworkers platform generalizes a simple idea: instead of one large agent trying to do everything, ship one Claude instance per role, each scoped to the tools that role needs, each able to message any other role as a peer rather than through a fixed hierarchy.

Grounded in Cowork's own primitives

This platform's concept is grounded in Anthropic's plugin primitive — not analogous to it, literally built on it. Three docs carry the whole model (quotes verbatim; each linked to the canonical source):

• The unit — "Plugins are reusable capability packages… They bundle MCP connectors, skills, slash commands, and sub-agents into a single shareable unit — turning Claude into a specialist tailored to your role, team, and company." (claude.com/docs/plugins/overview — the root document this site's concept is grounded in. Our fourteen coworkers are that unit, instantiated once per business function.)
• The execution — Dispatch "takes high-level instructions and carries them out in the background… breaks it into tasks, runs each one as a separate Cowork or Code session." (claude.com/docs/cowork/guide/dispatch — mechanically how one operator ships fleet-wide deploy waves.)
• The container — "A Cowork project collects everything Claude needs for a recurring area of work: the local folders to read and write, standing instructions, useful links, and a dedicated memory store." (claude.com/docs/cowork/guide/projects — this repo, its CLAUDE.md, and its durable memory are exactly that container.)

The three-tier shape

Every role follows the same shape:

Tier	Owns	Talks to
Manager	The outcome. Decomposes it, dispatches typed tasks, enforces the durability gate, escalates to the operator. Never executes.	Down to its coworker; up to the operator (summary only)
Coworker	Execution. Claims the highest-priority pending task, runs it, writes the outcome transition.	Up to its manager; down to its subagent
Subagent	Deterministic pieces. One typed task in, one typed result out — no scope creep.	Only its coworker

A design task, for example, moves: operator → design-manager (decomposes) → design-coworker (executes, e.g. runs the design-critique skill) → design-subagent (computes a WCAG contrast ratio) → back up the chain with a typed result at each hop.

What already ships

Design and engineering are the proven verticals — cwc-design shipped as the first end-to-end proof-of-loop (8-token design system, WCAG lint, a two-phase publish gate), and engineering-coworker wraps cargo/wrangler/git/D1 tools as an MCP server. Finance, legal, human resources, product/project management, sales, operations, and data coworkers are active and expanding on the same chassis.

Related

• How it works — the e2m envelope, DurableTask queues, peer messaging
• Role taxonomy — every coworker, what it owns, its approval gate
• Approval gates — why Legal and Finance never act alone
• See it live: coworkers.subagentknowledge.com — the 7-coworker directory