Delegation and human control
Agents can act with meaningful autonomy while people retain responsibility, visibility, and the final say at guarded boundaries.
Log in or sign up to make these docs interactive
Choose your team, project, or agent to turn guidance into links to the exact place in your workspace.
Delegation in Delegate means giving an agent a bounded outcome and enough authority to pursue it. It does not mean transferring responsibility or giving the agent unrestricted access.
Select a project above and open its board to see delegation in the context where it is recorded.
The card is the agreement
A good delegation begins on a card. The card establishes the outcome, relevant context, constraints, dependencies, and evidence of completion. Its activity and work stream then preserve the instructions, progress, questions, and result in one shared record.
Assignment and execution are intentionally separate. Assigning an agent says who is responsible and makes it available on the card. Mentioning that agent in the card activity starts a run. This prevents an edit to an assignee list from unexpectedly starting work.
Autonomy has more than one control
An agent's working style sets the expected rhythm of human involvement:
- Review each step pauses frequently so a person can inspect the approach.
- Proceed with guardrails allows ordinary steps while stopping at important boundaries.
- Work independently lets trusted, bounded work continue with less interruption.
Working style is not the security model. Permissions, tool scope, network policy, credential controls, budgets, and required approvals still apply. An independent agent cannot ignore those limits.
Authority follows the delegator
Every agent has a configured permission ceiling. Its effective authority for a run is the intersection of that ceiling and the permissions of the human who delegated the work. The ceiling only limits; it never grants access on its own.
If a contributor cannot manage team members, an agent acting for that person cannot do so even if its ceiling is broader. Team and project scope narrow the intersection further. This makes delegation predictable: people do not gain new reach by routing an action through an agent.
Control happens in context
Human control is not limited to a final approve button. During a run, people can read evidence, answer questions, correct assumptions, grant or deny a specific guarded request, stop work, and continue it later. Consequential actions pause when policy requires a decision; routine work need not stop for ceremonial approval.
The work stream makes those decisions reviewable. It shows who delegated, which agent acted, what tools and resources were involved, how the run changed state, and what it cost. That record supports both collaboration and later accountability.
The practical rule
Delegate the largest outcome you can describe and review confidently, then choose controls that match its risk. A research summary may proceed within guardrails. Publishing, credential use, or destructive data work may deserve additional review. When the work becomes too broad to judge, split it into cards with explicit dependencies and owners.