Agents learn Skills from repeated work

Task Machine can now review work an Agent has completed repeatedly and propose a reusable Skill when the pattern is clear. Future tasks can start with approved instructions instead of rebuilding the same approach from scratch. We built it because agents kept rebuilding the same approach from scratch on similar tasks.

Task Machine periodically reviews completed work for each Agent and recommends the smallest useful change supported by the evidence: assign an existing Skill to more Agents, improve a workspace-authored Skill, or create a new one.

When the evidence is not strong enough, nothing is proposed and no Inbox attention is required. Marketplace Skills can be assigned to additional Agents, but their content is never rewritten.

Set a separate policy for each kind of change

Every consequential Agent action now has its own review or direct policy. Task Machine learns from decisions for that exact action instead of pooling unrelated work, samples every tenth direct action for review, and proposes action-specific policy changes through the Inbox. Agent and Team Autonomy settings group these actions and provide bounded evidence details for each one.

When a learned-Skill action requires review or is sampled, its proposal arrives in the Inbox with the completed tasks that support it, a plain-language explanation, and any relevant limitations. New Skills show their complete instructions, while proposed updates show the existing and revised Markdown as a structured before-and-after diff. Eligible direct actions apply without creating Inbox attention.

For reviewed proposals, you can add or remove Agent recipients before approving. Approval publishes the Skill changes and Agent assignments together. Rejection leaves the current setup unchanged. The completed Inbox item keeps the reviewed evidence, content, and final recipients as a durable record of the decision.