cowork is an AI colleague. Not a chatbot that answers, not an autonomous agent running behind your back. A coworker you brief, review, and sign off on.
The agentic loop runs entirely in the browser. Your data, your keys, your integrations \u2014 never pooled on a shared server.
The most senior people in any organisation share one habit: they delegate. Not to avoid the work \u2014 to save their judgment for the decisions only they can make.The cowork thesis · 2026
Everyone wants adoption. Most organisations hit the same five walls. Delegation-first architecture was designed to clear them.
“Our adoption is stuck with the enthusiasts.”
Delegation is a UX everyone already understands. Brief a goal, approve an action. Skills pre-load your team’s methods so new users get expert behaviour from day one — no prompt engineering required.
“Nothing we do spans our systems.”
MCP makes cross-system work first-class. One delegation can pull from your CRM, draft in your email, update your project tracker, and post to chat — in one conversation, with approvals where they matter.
“We can’t standardise how we use AI. Who even owns it?”
Skills turn best practices into versioned, shareable units governed by OIDC groups. Ownership lives with the team that does the work — not a central AI committee that can’t keep up.
“The model landscape changes every month. How do we stay resilient?”
Pluggable LLM adapters: Anthropic, OpenAI, Google, Mistral, or self-hosted. Switch providers without migrating prompts, tools, or skills. Model-agnostic by design.
“Individual AI subscriptions are already getting expensive.”
One licence covers your team. Per-user cost tracking with quotas. Route light tasks to small models, reserve frontier models for the work that needs them. Pay for capacity, not seats.
Select a role to see a delegation run end-to-end \u2014 across your tools, pausing only where your judgment is needed.
Every piece of knowledge work decomposes into these nine capabilities. cowork handles the first eight so well you stop thinking about them \u2014 and hands the ninth back to you, every time.
Reads your docs, wikis, tickets, and inboxes. Stays grounded in what your org actually knows.
DelegableCarries context across sessions. Knows what you shipped last week without being reminded.
DelegableSurfaces patterns in data, draws connections between sources, turns noise into summaries.
DelegableRuns tools, edits files, fires API calls, closes tickets — the visible work of doing.
DelegableOrchestrates across services, teammates, and models. Keeps a plan moving in multiple places.
DelegableDrafts messages, docs, replies in your voice. Matches tone to audience and channel.
DelegableReads the room. Knows when a problem is actually a question about scope, priority, or politics.
DelegableAdjusts its approach when feedback arrives. Learns from corrections without being re-briefed.
DelegableThe call only you can make. What ships, what doesn\u2019t, what\u2019s worth the tradeoff. This one stays with you.
Yours, alwaysAnthropic, OpenAI, Google, Mistral, or self-hosted. Pick the right model per task; switch without migration.
Hand off several tasks at once. Notifications pull you back in when an agent needs a decision.
Connect any tool, build your own, swap providers. Zero lock-in — the agent discovers new capabilities as you add them.
Conversations encrypted in the browser before they touch the server. Envelope encryption, your keys.
Microsoft Entra ID, Google, or any OIDC provider. Group-based access to skills, servers, and quotas.
Install on any device. Read history offline. Background sync when you reconnect.
Early access · limited · no spam, occasional updates