I'm seeing two user experiences emerge around agents. One is at the org level. You standardize an agent for the company, and the absolute must is that it lives in your town square. Usually Slack, or wherever people already talk. You can start with a single machine. You probably should. But putting it somewhere people have to remember to visit feels like making the first mistake before you've even started.

We have an internally deployed agent now, primarily doing two things: knowledge work and tasks. It has context from our tools and answers surprisingly deep, compound questions. Then there's execution. Ask it to handle a customer-facing task or make a coding change. That's where one machine starts getting crowded. Port conflicts are a very unglamorous way to discover the limits of your agent strategy. We're moving toward fresh sandboxes for tasks, with an orchestrator managing work across machines. The conversation stays in one place. The work doesn't have to.

The other experience is at the builder level. Claude Code, Codex, Superset, Orca. Different tools, but once you're running several agents, you can end up maintaining several conversations too. What's this one doing? Did that one finish? Which thread had the decision? There is a point where the army of agents starts generating work for the person supposedly being helped.

I think a Grok Bot-style interface will become more common. Not necessarily that product, but the pattern: one primary agent you interact with, and however many agents it needs underneath. One conversation with something that gets shit done is a much more pleasant experience than being the message bus between five of them. The bottleneck continues to be human. Giving the human more threads to supervise doesn't make that go away.

My bet is that a lot of these tools converge on that interface. You tell one agent what you need, clarify things with it, and review the result. It handles the delegation. In hindsight, this feels fairly inevitable. We got excited about running multiple agents, then discovered we also had to talk to all of them. Of course we'd eventually want one conversation again.

Everybody is becoming a manager. Fair enough. But high time for everybody to build, too. The space is still very new, and we're already spending this much energy figuring out how to keep up with the work agents can do. The thought I keep coming back to: if the trajectory holds, this is the least capable AI we'll ever use.