Agent UIs people can trust
Streaming output, every tool call visible, approve-or-reject for risky actions, and an interrupt that really stops the loop.
AI · MCP
Agents, from
the interface out
An agent is a loop: the model asks for a tool, the app runs it, the result goes back, until the model says it's done. The engineering is in making that loop visible, safe and fast for a person.
Fig. 06 — one tool-using turn
UserAgent UIModelMCP clientMCP server
- 01 “Which transactions failed today, and why?”from User to Agent UI
- 02 messages + 2 tool definitionsfrom Agent UI to Model
- 03 stop_reason: tool_use · transactions.searchfrom Model to Agent UI
- 04 run toolfrom Agent UI to MCP client
- 05 tools/call {status: failed, date: today}from MCP client to MCP server
- 06 12 recordsfrom MCP server to MCP client
- 07 tool_result (matching id)from Agent UI to Model
- 08 stop_reason: end_turnfrom Model to Agent UI
- 09 streamed summary, tool call visiblefrom Agent UI to User
Where I focus
3 layersMCP servers with real contracts
Small, well-described tools with strict input schemas. The same discipline I bring to component APIs, applied to what a model can call.
Loop control
Explicit branching on stop_reason, tool results returned with matching ids and errors, and iteration caps so a confused agent fails loudly.
Written up in more depth: Multi-agent orchestration ↗ · Agentic loops ↗ · All writing →