Skip to content
Agents for Humanity
Public draft. All results are preliminary desk assessments against criteria v1.0, updated 27 Sept 2026. No agent has been certified yet. How we assess
Autonomous agenthostedProprietary

Microsoft Copilot Cowork

Microsoft · microsoft.com/en-us/copilot/features/cowork

Cowork is a cloud-hosted, proprietary agent that runs delegated multi-step tasks on Microsoft models and Anthropic Claude, with approval prompts, pause/cancel controls, owner-written custom instructions and skills in the open SKILL.md format. The agent cannot be exported or run outside Microsoft's service, its system instructions and runtime are closed, task logs are kept for undisclosed periods, and access requires a paid Microsoft 365 tier.

Strengths
  • Pause, hard pause and cancel for running tasks; scheduled tasks can be paused or deleted
  • Approval prompts with previews and risk levels before sends, posts and meetings
  • Owner-editable custom instructions (up to 20 KB) applied to every task
  • Custom skills and plugins use the open SKILL.md and MCP formats
  • Local-browser tasks keep cookies and credentials on the owner's device
  • Choice among several Microsoft-offered GPT and Claude models
Gaps
  • No exportable or self-runnable agent; runs only in Microsoft's cloud
  • System instructions, built-in skills and runtime configuration are hidden
  • Task logs, screenshots and outputs kept for undisclosed periods; human review
  • No owner-facing tamper-evident audit log for personal accounts
  • Custom skills documented only for work accounts; personal edition is paid preview
  • No local or arbitrary model option
Evidence

All 34 findings

Copilot Cowork for personal Microsoft accounts (Preview, Microsoft 365 Premium or Pro), which replaces consumer Copilot Tasks, as documented on 2026-09-27. Shared mechanics are taken from Microsoft Learn's Cowork docs, which are written mainly for work accounts; the work/school GA edition, tenant admin and Purview controls are not assessed.

Portable · Can you leave, and take the whole agent with you?

0%

Transparent · Can you see everything the agent is, with ordinary tools?

25%

Auditable · Can you reconstruct exactly what the agent did?

10%

Verifiable · Can you prove the agent runs what it claims?

0%
  • V1
    Open, reproducible runtime

    The Cowork runtime is proprietary, closed-source Microsoft service code.

    Fail
  • V2
    Active config is inspectable config

    The owner can inspect custom instructions, installed plugins, skills and the chosen model, but not the full server-side configuration, system prompt or built-in policies Cowork runs with.

    Fail
  • V3
    Attributable messages

    Emails and messages Cowork sends are authorized through the owner's Microsoft account, and recipients cannot verify that they came from this agent.

    Fail
  • V4
    Independently checkable record

    No action record is available to owners whose integrity can be checked with open tools that do not depend on Microsoft.

    Fail
  • V5
    Comparable state

    The owner cannot obtain Cowork's full state, so cannot verify whether it changed between two points in time.

    Fail

Modifiable · Can you change anything, without asking?

25%

Controllable · Is your word final?

50%
Something wrong or out of date?

Vendors and the public can dispute any finding with evidence. Disputes and their resolutions are published.

Dispute a finding