Drawing Agents as a tree is cool, but in production you probably just want to see tasks

Flo posted a demo of an Agent Orchestration interface on X. Multi-Agent runs are drawn as a tree, and each call expands its trace below, updating in real time.

Under this post, the first things to appear were usage questions; the architecture discussion had not started yet. Deethin.Design asked what tool was used for orchestration. Harley Lewis Foote then raised the bar: every call needs a trace, and tool calls and MCP responses also need granularity—you can't just provide surface-level metrics.

Jason Kneen's reply was more direct: MY EYES!!! is there dark mode? ;) Flo said yes and gave the affectUI Agent Orchestration screen.

The link points to affectUI's Agent Orchestration screen. The site positions itself as Agent-ready React UI for AI products, one component per file, no dependencies. The focus here leans more toward interfaces for AI products; the orchestration engine itself is not the star of this page. It has components like Agent Tree, Trace Waterfall, Tool Call Card, and Approval Gates, and the names already explain what it wants to solve: lay out the agent's runtime process for people to see.

JJdoesTech's reply offered another perspective. He said the UI was beautiful, but he wasn't sure it solved human attention. He himself had completely shifted in the UI to tracking tasks, not tracking agents. His exact words were chief of staff vibes and it's good. This judgment is worth pausing on. Multi-Agent systems are easiest to turn into topology diagrams, with nodes connecting to nodes, looking very busy. Once they enter daily use, users may only want to know which thing is stuck, who should act next, and whether approval is needed. The Agent tree is good for debugging and demos; a task list is better for allocating attention. The two don't conflict, but their priorities differ.

Another detail is granularity. Harley wants not only who called whom among agents, but also tool calls and MCP responses. MCP can be understood as a protocol for models to connect to external tools and data. For developers, surface-level metrics are like a dashboard; tool calls and responses are the actual scene of the incident. With one less layer of trace, troubleshooting requires one more layer of guessing.

affectUI made dark mode into a joke in the replies, and that also shows a reality of this kind of product: the demo video makes people say wow first, but when it's actually used, the first question is whether it strains the eyes. A beautiful UI is the entry ticket; trace granularity, task views, permissions, and approvals are the daily reality. The interface design for multi-Agent orchestration may start with a tree, but ultimately lands on how attention is allocated.