7 Agents Running, 1 Waiting for You: What a Multi-Agent Interface Really Should Show

Flo posted a component called Agent Fleet. The problem it solves is very specific: when multiple agents and their sub-agents are running at the same time, how does a person know what to look at right now.

Looking at the screenshot, the top row shows numbers: 4 agents, 14 sub-agents, 7 running, 1 needs your input, total cost $3.84. On the left are several main roles: Launch lead is using Opus 5.5 to ship a new pricing page, due Friday, 0% progress, needs your input, cost $1.65, elapsed time 14 minutes; Support lead is using Sonnet 5.5 to answer 48 open tickets in the queue, 50% progress, cost $0.79; Release manager is using GPT-6 Sol to release version 2.4 to the app store, 33% progress, one failure, cost $0.57; Market analyst is using Gemini 3.8 Pro to write the weekly competitor report, completed, cost $0.82. On the right are subtasks: Researcher compares prices from five companies, completed; Writer drafts copy for three platforms, needs your input; Coder builds the page and connects the code, running; Tester runs checkout tests, running.

The most noteworthy part of this interface is the “needs you” marker. The model names and the cents in the costs are moved to the back. In a reply, Flo said people need to see where action is required. One commenter said that with seven agents running and one waiting for you, that little badge might be the most important thing on the screen. Another commenter mentioned that the “Needs you” marker with an inline Answer button makes waiting agents hard to ignore.

But commenters also pushed the problem one level deeper. John Rood said that if clicking “Answer” still requires opening the agent to read why it stopped, then it is just a notification with an extra step; that row should directly surface the pending question and what you are choosing between two answers. Aleksandar Janca's suggestion is even stronger: next to the status, it should show what each agent is allowed to touch. Knowing that it is running matters less than knowing whether it can write to production. Ari asked whether stopping a parent agent also stops its sub-agents, and whether external actions already completed would be shown. If the parent stops but the child keeps running, that is an expensive accident. Koroo asked whether cards waiting for you to handle them jump to the top or stay in place within the tree.

These comments point to the same thing: a multi-agent management interface cannot just be a monitoring dashboard that says “everyone is busy.” It has to answer three questions: who is stuck, why they are stuck, and what a person can do here. Cost and elapsed time are outcomes; intervention points are the control surface. Flo replied to some of these suggestions that he would try them. The component is at https://affectui.com/components/agent-fleet.