Designing a workspace where people and agents share the work

Once a team has agents that can take real actions, the question is no longer whether to use them but how far to trust them. Too little trust and the agent saves no time. Too much and a single bad action can cost more than everything it saved.

Think in levels, not a switch

A useful way to frame it is to give each type of task its own level of autonomy:

• Suggest: the agent drafts, a person decides.
• Act with review: the agent acts, and a person checks before it takes effect.
• Act and report: the agent acts and leaves a clear record for later review.
• Act alone: kept for tasks where the agent has a proven record and mistakes are cheap to undo.

Earn each level with evidence

A task should move up a level only when there is evidence the agent handles it well: a record of reviewed outputs that met the standard, including the difficult cases. It should move back down when that record slips. This turns trust into something visible and adjustable, rather than a one-time decision.

Autonomy is something an agent earns on a specific task, not a setting it is given everywhere.

Make review quick enough to actually happen

Review only works if people do it. That means short, focused checks against a clear standard, not open-ended reading. A reviewer should be able to answer “does this meet the bar, yes or no, and why?” in minutes. When review is that clear, it can be done consistently and its results can be trusted.

Keep the record

Every action an agent takes should leave a trail: what it did, what it based the decision on, and who approved it. The record is what lets a team find the cause of a mistake, fix the standard and prevent it next time. Without it, every failure is a mystery.

Signs a task is ready for more autonomy

• Reviewed outputs have met the standard consistently, including on hard cases, not just easy ones.
• The cost of a mistake is low, or the mistake is easy to catch and undo.
• The agent reliably flags its own uncertainty instead of guessing.
• Reviewers agree with each other about what counts as a pass, so the record means something.

If any of these is missing, the safer move is to keep the task where it is and keep collecting evidence.

The human role

In this model, people are not supervising every keystroke. They set standards, handle the cases agents escalate, and decide where autonomy is safe. It is less work by volume, but it carries most of the responsibility, and it needs people who know the domain.