You ask an AI agent to make a change to a website. It reads files, writes code and runs checks while you work on something else. When it returns, you have a decision to make: accept the result, ask for a change, or send it back to investigate.
That is a different kind of work from changing every line yourself. The person has not disappeared. The job has moved: from directly operating the tool to setting a direction, watching what happens and deciding when to intervene.
The change reaches beyond programming. An agent might gather material for a report or prepare a publication. This publication runs on Papyrus, a system designed to support AI-assisted newsroom work alongside human editorial direction. We have a connection to that example: Papyrus also powers Machines on Machines. A conventional publishing editor already offers automation and editorial controls; the difference here is the possibility of delegating more of the work between decisions, not a claim that other systems have no automation or that people must give up control.
The steering wheel needs a view
The idea has an older relative in cybernetics, the study of control and communication in machines and living things. A thermostat offers a simple feedback loop: it senses a temperature, compares it with a setting and adjusts the heat. The result of the adjustment becomes the next measurement.
A person working with an agent can participate in a less tidy version of that loop: set a goal, observe the agent's work, correct it and see what happens next. But a loop works only if information can travel back. If all you see is a polished final answer, you may have little idea what the agent tried, what it overlooked or when it made an important decision.
Researchers are beginning to treat that visibility as a design problem. A November 2026 preprint about a coding tool called Vibetrace describes an interface that lets users inspect a record of an agent's actions, return to an earlier point and give a targeted correction. The authors report promising results in tests and user studies. It is a research prototype, not proof that every coding agent can be reliably supervised this way. 1
Another project, MUSE, takes on a similar problem in data science. It organizes an agent's execution history at different levels of detail and lets people ask about or revise particular steps. In a study involving 15 participants, its researchers reported better task efficiency and greater confidence in understanding and steering the workflow. That small study is a useful signal, not a guarantee that the approach will work everywhere. 2
When should the machine ask?
Visibility alone is not enough if a person has to stare at every action. The point of an agent is partly that it can handle some work without constant direction. But the further it goes, the harder it can become for someone to know when they need to step in.
Stanford researchers have explored ways for an agent to learn when to request human guidance, including before potentially risky actions. Their experiments offer a possible answer to the timing problem: instead of asking a person to approve everything or trusting the system to proceed indefinitely, design the handoff around moments when guidance matters. It remains research, not a solved rule for all agents. 3
The balance already varies by task. Anthropic's 2026 analysis of activity on its own services distinguishes delegation from several more interactive patterns, including work in which a person supplies feedback. Those records describe one company's users and products; they do not establish how every workplace uses AI. They do show why a single phrase like “human in the loop” can hide very different arrangements. 4
Watching is work, too
A person can be nominally responsible while lacking the time or information to exercise that responsibility. Researchers have argued that agent systems can make meaningful oversight harder, especially if their users become less practiced at checking the underlying work. That is a warning and a research argument, not proof that every agent makes its operator less capable. 5
The design questions are concrete. Can the person see what the agent used? Can they stop or redirect it before an action matters? Can they examine a meaningful sample of the work without drowning in logs? Can they learn from a failure and change the next attempt?
Stepping back can be valuable. It lets people spend less time on repetitive operations and more time deciding what they want to accomplish. But distance also changes what they can notice. The promise of a good feedback loop is not that the human holds every control at once. It is that the system gives them a useful view, a timely way to act and a chance to see the result.
References
References
- From Watching to Steering: Enabling Human-Coding Agent Collaboration through Execution Traces. (n.d.). https://arxiv.org/abs/2611.04466
- MUSE: An Interactive Meta-Agent for Understanding and Steering LLM-powered Data Science Systems. (n.d.). https://www.microsoft.com/en-us/research/publication/muse-an-interactive-meta-agent-for-understanding-and-steering-llm-powered-data-science-systems/
- A blueprint for keeping humans in control of AI. (n.d.). https://news.stanford.edu/stories/2026/09/ai-human-control-frameworks
- Anthropic Economic Index report: Cadences. (n.d.). https://www.anthropic.com/research/economic-index-june-2026-report
- AI Agents Push Humans Out of the Loop. (n.d.). https://arxiv.org/abs/2608.23642