Assistant–UserProxy Pair

Two-role micro-workflow where a tool-using Assistant acts and a UserProxy keeps it on the human goal, corrects drift, and declares completion.

An Assistant–UserProxy pair is a minimal multi-agent architecture used by frameworks such as AutoGen:

• Assistant — a standard conversational LLM agent with a ReAct-style loop and background tools.
• UserProxy — stands in for the human; system-messaged to pursue the stated goal, act as a corrective force (steer, recommend), and eventually declare the goal accomplished. Often tool-light or tool-less relative to the Assistant.

The pair itself is a small LLM workflow: dialogue between roles replaces a single agent that must both act and self-correct. System messages and tool allocation are the harness levers — a CodeAssistant with file/test tools and a goal-bearing UserProxy with no tools will converge or loop depending on those choices. Scale the same idea with a group chat manager or hierarchical director that delegates among several role-specialized agents (LLM-driven workflow routing).

---

Sources

• Prompt Engineering for LLMs (Berryman), ch. 9