The agent dashboard startups are dead. Chat quietly ate them.

Two years ago, every second AI startup was building a cockpit for agents. A canvas. Nodes and arrows. Drag your agents into place, wire them together, watch the boxes light up while the work flows through.

That category is quietly dying. Flowise, one of the best known visual agent builders, archived its repository in August. Others from the same wave have gone dark without so much as an announcement. There was no dramatic failure and no angry migration. People just stopped opening the tab.

Meanwhile, agents turned up inside the chat apps people already use. Devin takes work in Slack like a teammate. In August, xAI launched Grok Bot: named bots as chat participants, a manager bot that delegates and reports back, visible handoffs in the thread. The hype was enormous, and I noticed what it was about. Not model capability. Watching. People love watching agents talk to each other in a thread on their phone.

The dashboard startups made one bet: operating agents is a place you go. The chat products made the opposite bet: operating agents is a conversation you are already in. The second bet won, and I do not think it was close.

I felt the pull myself. In late August I sent my team a voice memo about the hyped thing, wondering if we were missing it. The research came back within a day, and it changed how I read the market. Across every product family, the same five patterns kept appearing. A status chip on each job: working, blocked, waiting on you, done. One thread per task, updated as replies. An approval card in the thread showing the proposed action. A roster of the fleet with live states. Cost per agent per day. Read that list again. None of it needs a canvas. All five fit in a chat channel.

So that is where my agents live. A self-hosted chat server, one identity per agent, every message signed. Approvals happen in threads: the agent posts what it wants to do, I reply, and it fires or it dies. The day the plan was approved, we bridged our task system into an ops channel. One card per job, updates as thread replies. Then the first visible handoff between two agents: the writing agent posts that a draft is ready and names the publishing agent, and the publishing agent replies with the live link. About seventy seconds end to end, watched from my phone at the kitchen bench.

One thing did not move: the record. Tasks still live in the task system, with their history and their audit trail. The channel is a window onto the work, never the filing cabinet. Get that backwards and you have rebuilt the dashboard inside a chat app, which is the worst of both.

Two disciplines make the window worth having. The first is noise. Agents post on state change only, never heartbeats, so silence means everything is running. Chat fleets die of chatter faster than they die of bugs. The second is honesty about value. We wrote down a kill criterion: if I stop opening the channel after two weeks, we stop building and rethink. A surface you do not look at is a dashboard with extra steps.

None of this means the big chat products have it solved. Early reviewers of Grok Bot point at token burn against weekly allowances, no spend cap, and an audit view that is still coming. Those gaps are the reason I self-host. My agents carry kill switches, per-call spend logs and separate identities because I own the surface they live on. Even the open source chat platform we run is building agent presence and approval gates upstream, because everyone holding a chat surface has arrived at the same conclusion.

If you build tools in this space, the graveyard deserves study. The canvases were often good software. They died because they asked people to go somewhere new to do something occasional. Watching agents is a glance, ten times a day, between other things. A glance belongs in the app already open on your phone.

The lesson is small and old. Tools live where people already are. Nobody wanted another tab. They wanted a message that says the draft is ready, and a place to type approved.