Independent briefing Work & economy
Agency and agents
A practical argument for AI agents that increase human agency instead of quietly taking it away.
Independent briefing of One Useful Thing (2026). Read the original for its full evidence, context, and wording.
Start here
Three questions this article answers
What makes an AI system an agent?
It does more than produce an answer: it plans, uses tools, coordinates steps, and acts toward a goal with some independence.
Why is human agency the right design test?
Because speed is not the only outcome that matters. A useful system should expand people’s ability to choose, understand, and intervene.
What is the “Twilight Factory” alternative?
A workplace where agents handle routine coordination but deliberately call people in for approval, expertise, diverse judgment, and the decisions people actually value.
Go deeper
Three related questions it doesn’t answer—matched to established papers
Which workers gain the most from AI assistance in real workplaces?
A field study of 5,179 support agents found the largest productivity gains among newer and lower-skilled workers.
Paper answering this questionGenerative AI at WorkWhy does the same model help on some tasks and hurt on others?
A widely discussed field experiment with 758 consultants shows that AI’s effects depend sharply on whether a task sits inside the model’s capability frontier.
Paper answering this questionNavigating the Jagged Technological FrontierHow should work be divided between people and automation?
A foundational human-factors paper explains why appropriate reliance matters more than maximum automation.
Paper answering this questionHumans and Automation: Use, Misuse, Disuse, AbuseThe useful summary
What to take away
The article reframes the agent debate around a human outcome: agency. Agents can coordinate tools, pursue goals, and take actions, but autonomy is not automatically progress. A system that completes more work while making its operator less informed or less able to intervene may be efficient and still be badly designed.
Its strongest idea is the Twilight Factory: an organization where software handles the dull connective tissue of work, yet actively looks for moments when human taste, accountability, context, or approval would improve the result. The point is not to preserve every old task. It is to keep people attached to meaningful decisions and to make escalation a feature rather than a failure.
Final conclusion
The three answers, brought together
An agent is an AI system that can plan and act, but its value should be judged by the agency it creates for people. The better future is not a lights-out organization. It is a “Twilight Factory” where machines absorb coordination and people remain visible, informed, and decisive.