I don't see any good orchestration system for AI agents

A developer argues that no good orchestration system exists for AI agents, noting that most current solutions are superficial and that practitioners resort to primitive split-terminal setups. The author highlights the lack of environment isolation, workspace management, and human-in-the-loop control, and calls for a system that treats code as important but secondary, allowing delegation with easy intervention.

A couple of weeks ago I saw a guy in the office with six instances of claude code, each in a different window. I've seen a lot of attempts at orchestrating agents. And there are some viral repos which show a lot of agents like videogame characters, and you can touch any one of them and then chat with that agent. But none of these solutions are actually useful; it's all a facade. Because at the end of the day, people keep going back to the split-terminal setup. There's no real solution for the environment problem. I don't feel comfortable running claude --dangerously-skip-permissions on the machine I have all the data. I would like to run each agent in a docker container. I'm sure there are a lot of images for claude code and even libraries that trigger tasks and whatnot. But I don't see any commonly adopted solution. I feel like a lot of people are talking about agentic workflows and a lot of agents running simultaneously, but the ones that are actually running agents in parallel for doing actual work are doing it in the most primitive way: just a handful of terminals. Which brings a lot of problems. Because what about the workspaces? I can set up worktrees for each, but then what if I want to review their work? One of the main problems I have with a lot of the solutions in the space is that they assume it will just work. There is no easy way for controlling an agent and stepping in. Opus 4.8 is amazing, but there are times when it's a lot easier and cheaper to just look at the code and change that one variable. There seem to be these two big groups currently in tech: - The ones that treat code like a black box, and never touch or see the code no matter what - The ones that use AI, but do so in a very controlled or "conservative" way: a couple of agents running, but watching them, watching the code. They don't really trust AI. The first group loses a lot when hard problems appear, but the second group is lagging behind the first. And yes I do think that blindly telling AI things makes for a very fragile architecture that becomes an unmaintainable mess in a few months. But companies or clients will judge you on your work, and will value how much you did over what problems you avoided. So, for better or worse, the second group is lagging behind and seen as "unproductive". What seems to be the winning move is to delegate work to the agents, while being willing to go into the code and read it and "get your hands dirty". But I don't see a library for that. That is, for delegating control, while at the same time allowing you to step in and check what's happening. I tried zed editor, and it's quite nice, but it still has this "code-first" approach. It also does not have support for VS Code plugins. I think that vibe coders are starting to get right most of the time that "you don't need to know programming" to build things, but there are crucial moments where stepping in is the best, and they pay very hard the price for not doing that. The code is still relevant, but not so much that you have to watch it closely every time. If you design it correctly, and the AI understands you, it will probably write good code. Even Linus agrees that it is reaching pretty good levels. I think the code is starting to be more secondary. Which also opens up a possibility that only Cursor and Replit seem to be exploring: starting coding on the smartphone. But that is another topic. The point is, there is no solution that makes code an important but not primary actor. Current solutions or libraries make the agent a part of the workspace, and not the other way around, which is starting to seem the most sensible approach.