0083 — Cowork mode: a prepackaged knowledge-worker archetype

• Status: accepted
• Date: 2026-07-16
• Deciders: Josh, agent

1. Context & problem statement

Claude Cowork is Anthropic's agentic workspace for non-coders — verified by teardown to be Claude Code headless in a sandbox VM. By mid-2026 it ships on every paid plan, and Anthropic's own 1.2M-session study shows who lives there: business operations (33.4%) and content creation (16.4%) dominate; software development is 8.7%. The deliverables are files — spreadsheets, documents, decks, PDFs — produced by four default office-document skills, plus a schedule skill (distill the session into a self-contained prompt, run it on a cadence), a consolidate-memory skill (merge/prune memory files), and a guided setup-cowork onboarding (role → role-matched plugin packs → the connectors those packs declare → try a skill → a saved writing-voice profile).

Those users hit verified ceilings that protoAgent already exceeds: cadence limited to hourly/daily/weekly (no cron), cloud scheduled tasks that cannot touch local folders, memory siloed per-surface and projects-only, no model choice, usage limits that "burn faster than chat", and — since remote sessions became the default — local files processed on Anthropic's servers. The self-hosted pitch writes itself: same workflow, data never leaves the box.

But protoAgent has no landing pad for these users. Every archetype in the new-agent picker that ships tools is developer-shaped (coder / PM / research); generic-assistant.md exists as an orphaned soul preset with no catalog row and no tools. A Cowork graduate today would have to hand-assemble plugins, skills, folder fencing, and a persona before protoAgent resembles what they left.

2. Decision drivers

• One-click: the existing archetype seams (ADR 0042) already deliver SOUL + bundle plugins + config defaults on both the first-run SetupWizard and the fleet new-agent panel — v1 must not require new core seams.
• Fidelity where it earns trust (file deliverables, folder consent, never-delete-without-asking, schedule + memory hygiene), divergence where we're strictly better (cron, local scheduled work, unified memory, model choice, plugins/fleet/A2A as the extension story instead of ZIP uploads).
• Legal cleanliness: Anthropic's document skills are all-rights-reserved (no redistribution, no derivatives, no extraction from the Services) even though the anthropics/skills repo is public.

3. Considered options

1. Catalog archetype + bundle (chosen) — a cowork soul preset + catalog row pointing at a cowork-stack bundle repo; skills ship inside a cowork-plugin listed by the bundle.
2. Fork-style "cowork edition" of protoAgent — rejected: modes are configuration, not forks; the operator-fork contract reserves forks for domain rewrites.
3. Core-seam-first (bundle mcp: field, env seeding, approval tiers before shipping) — rejected for v1: nothing in the package strictly needs them, and each is independently useful later.

4. Decision

• D1 — Ship as data, zero core seams. config/soul-presets/cowork.md (persona written from the usage data: ops/content copilot, file deliverables, folder-fenced, memory-forward, never delete without asking) plus one archetype-catalog.json row (id: cowork, bundle:https://github.com/protoLabsAI/cowork-stack, soul_preset: cowork), inserted before custom.
• D2 — cowork-stack bundle composition. cowork-plugin (the skill pack) + artifact + notes + execute_code (builtins; the document skills need a runtime — the same reason Cowork carries a VM) + google-plugin (Gmail/Calendar/Drive ≈ the connector layer) + claude-bridge (imports the user's existing Claude Code/Cowork state). Curated enabled list + config defaults (operator fs tools fenced to user-chosen folders, scheduler on).
• D3 — Original document skills, not Anthropic's. cowork-plugin ships clean-room docx/xlsx/pptx/pdf skills over the standard Python libraries (python-docx, openpyxl, python-pptx, pypdf), declared via requires_pip. Vendoring Anthropic's skills is prohibited by their license; that finding is recorded here so nobody re-litigates it.
• D4 — Migration respects the same license. The claude-bridge skill importer must skip creatorType: "anthropic" skills and migrate only user-authored ones (including my-writing-style).
• D5 — Deferred seams, filed as issues. A bundle mcp: field, a read-write-no-delete fence mode, per-action approval tiers (ties into the ADR 0071 consent-acks follow-up), and env/secrets seeding from archetypes.

5. Consequences

• Good: Cowork's entire prepackaged experience becomes reproducible self-hosted in one click, and the archetype doubles as the template for future "modes" (a role pack is just another bundle + preset).
• Trade-off: until D5's seams exist, Google OAuth and any MCP servers are wired up in the console after install rather than seeded by the archetype — the onboarding skill walks the user through it, mirroring setup-cowork.
• Risk: the doc skills are original prompts over commodity libraries; parity with Anthropic's tuned skills is asymptotic, not guaranteed. Evals (ADR 0012) can close the gap iteratively.