> 📥 The full comparison is available as a PDF (Danish, 3 pages). Download it at the top of the page — three columns, five routines and all sources in one place.
Three paths from coding agent to work agent
OpenAI builds Work on Codex technology. Anthropic uses the same agentic architecture as Claude Code. Microsoft has integrated the technology behind Claude Cowork into Microsoft 365 Copilot. Three different workspaces, one direction: you describe the outcome, the agent plans and does the work.
The 2026 timeline is short and dense:
- January 12 — Anthropic: Claude Cowork launches as "Claude Code for the rest of your work".
- June 16 — Microsoft: Copilot Cowork becomes generally available globally.
- July 9 — OpenAI: ChatGPT Work launches with Codex technology built in.
01 · Folder-first: Claude Cowork
The proven specialist.
- Engine: The same agentic architecture as Claude Code, without the terminal.
- Workspace: Remote sandbox in beta. Local files, browser and computer via the desktop app. Projects hold files, rules and memory.
- Delivers: Formatted documents, organised files and research. Plugins, skills, subagents and schedules extend the way of working.
- Best when: Work starts in folders and templates, and you want to steer long, multi-step tasks closely.
- Watch out: Agent security remains an active area of development. Cowork consumes far more capacity than regular chat.
- What you can review afterwards: Plan, progress and actions within the session. Cowork is currently not covered by Anthropic's Compliance API; Team/Enterprise can use OpenTelemetry.
02 · Tenant-native: Copilot Cowork
The enterprise-integrated path.
- Engine: The technology behind Claude Cowork, integrated into Copilot. Multi-model and connected to Work IQ.
- Workspace: Protected cloud sandbox inside Microsoft 365 with Outlook, Teams, files, Fabric, Dynamics and plugins.
- Delivers: Documents, Excel analyses, slides, emails, calendar changes and updates across M365.
- Best when: Microsoft 365 is your system of record, and existing identity, permissions and compliance must carry over.
- Availability now: GA since June 16. Requires an M365 Copilot licence, usage-based billing and admin activation.
- What you can review afterwards: Actions and output are auditable. Sensitive steps, such as sending mail or updating a record, stop for approval.
03 · Broad toolbox: ChatGPT Work
The hybrid work agent.
- Engine: GPT-5.6 with Codex technology built in.
- Workspace: Cloud on web and mobile. Desktop can work with local files, apps and a built-in browser.
- Delivers: Documents, spreadsheets, slides, reports and Sites. Plugins and Scheduled Tasks tie workflows together.
- Best when: You work across many tools and want to combine local files, cloud apps, the web and finished artefacts.
- Watch out: Rollout is gradual, and features vary between web, mobile and desktop. PowerPoint is not part of the desktop Work flow at launch.
- What you can review afterwards: Progress and approvals within the task. Enterprise/Edu can use the Compliance API for insight into Work conversations and actions.
The functional overlap is now hard to miss
All three can:
- plan multi-step tasks
- work across files and apps
- research and use a browser
- produce finished artefacts
- keep going while you do something else
The decisive question is control
Once the agent has access to mail, CRM, files and browser, you must be able to reconstruct the sequence: which sources it used, which files it changed, which actions it proposed or performed — and who approved them.
> My take: Choose the platform based on your system of record. Choose the control model before you decide how autonomous the agent may be.
Before you hand over the keys
Five routines separate an impressive demo from a work agent you can actually stand behind.
1. Limit access. Start with few folders, few apps and work identities. Use read-only where the agent only needs to understand.
2. Set approval thresholds. Mail, CRM updates, sharing, deletion and payments require an explicit human gate.
3. Keep a usable trail. Plan, sources, changed files, external actions, approvals and errors must be findable again.
4. Build in rollback. Version history, backups and one canonical document make mistakes reversible. Test restore before scaling.
5. Put accountability on a human. Every agent and every workflow needs a named owner, a review criterion and a fixed rhythm for access and log reviews.
> The principle: Autonomy without traceability is debt. Start with the traceability you will wish you had the day something goes wrong.
Sources and method
- [1] OpenAI, July 9, 2026: ChatGPT is now a partner for your most ambitious work.
- [2] OpenAI Help: Creating and editing files with ChatGPT Work.
- [3] Anthropic Help, updated July 2026: Get started with Claude Cowork.
- [4] Anthropic, July 7, 2026: How people are using Claude Cowork.
- [5] Microsoft, June 16, 2026: Copilot Cowork is now generally available.
- [6] Microsoft 365: Copilot Cowork product page and the March 9, 2026 launch.
> ⚠️ Disclaimer. General practical guidance — not legal or compliance advice. Always involve your own compliance or legal function and your DPO, and verify dates and status against primary sources. Status as of July 2026.
📬 Subscribe to the newsletter "AI, Built Human" on Substack — weekly insights on AI in practice.
Stefano Vincenti · GenAI strategist and architect · External lecturer, IT University of Copenhagen & DIS Copenhagen · Cofounder & CTO BotTellMe · Partner, TryZone · aitrainer.dk