← Back to concept explainers

Concept Fable

User Proxy Agent

Conclusion first:User Proxy Agent is a role in multi-agent systems that represents the user, commonly seen in AutoGen. It can provide human input, approve actions, execute controlled code on the user's behalf, or hand questions back to a real person. Its core value is establishing a clear interface between human and agent collaboration: which actions run automatically, which require confirmation, which information can be exposed, and which executions can only happen in a sandbox.

User Proxy Agent · Multi-agent role

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

The person who sits at the council table on behalf of the master

When the city was building a new bridge, the engineers, accountants, and inspectors met every day. The chamber of commerce president who actually put up the money was busy and only signed at the end. In the early meetings no one represented his preferences, so proposals were often rejected halfway through.

The engineer said it needed to be sturdier, the accountant said it needed to be cheaper, the inspector said it needed to be safer. Everyone made sense, yet no one could answer whether the chairman would actually approve more budget, which risks were acceptable, and which decisions had to wait for him personally.

Someone proposed moving all the letters the chairman had written in the past into the meeting room. Everyone looked through many old letters, yet still could not accurately guess how to weigh today's matter.

Later, the chairman sent an agent to sit at the table. The agent could not draw bridges or calculate beams for the engineers, but he knew which questions could be confirmed on the spot and which had to pause for the chairman's approval.

At the first meeting, the engineer wanted to try a dangerous lifting operation on site. The agent immediately stopped it, saying this required explicit permission. The accountant wanted to check the chamber of commerce ledgers, but he gave only a budget range and did not open all the books.

Sometimes the agent also takes action. It brings back the additional requirements from the chair, or runs a short test in a safe sandbox and hands the results to the other masters.

Meetings began to go smoothly. Others understood that the agent is neither an all-powerful master nor an ordinary observer. It is the seat for human intent, permissions, and confirmation mechanisms within the collaboration system.

When the bridge was finally built, the chairman found that he did not need to attend every detailed discussion, nor had he handed decision-making entirely to the machine. At key nodes, someone guarded the gate for him, and when necessary brought the issue back before him. The agent said: a representative participating in a system is not about imitating all human abilities, but about putting human boundaries, permissions, and input in the right places. Later, other projects also invited agents to the table. Everyone first confirmed which preferences he could represent, which materials he could call on, and which boundaries required checking back with the person himself, so the meeting would not mistake convenience for authorization. The agent also left a record of the basis for decisions, making it easy for the chairman to review afterward. If his judgment drifted from the chairman's true preferences, the boundaries had to be corrected before the next meeting.

Reveal

This story is about: user proxy Agent

User Proxy Agent is a role in multi-agent systems that represents the user, commonly seen in AutoGen. It can provide human input, approve actions, execute controlled code on the user's behalf, or hand questions back to a real person. Its core value is establishing a clear interface between human and agent collaboration: which actions run automatically, which require confirmation, which information can be exposed, and which executions can only happen in a sandbox.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Chamber president: real user or business owner
  • Proxy: User Proxy Agent
  • Does not draw the bridge but can confirm boundaries: the agent does not reason through all content on behalf of the expert Agent, but manages user intent and permissions
  • A dangerous lift was stopped: high-risk actions requiring human approval
  • Give only a budget range: minimal necessary information exposure
  • Secure sandbox testing: controlled code execution or tool execution
  • Bring the problem back to the chair: human-in-the-loop
  • Final insight: User Proxy Agent is the controlled interface for human participation in multi-Agent systems

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

AutoGenhuman-in-the-loopassistant agentapprovalcode execution
京ICP备2026014272号