Coding agents were supposed to save us time. Instead we somehow became middle management for robots. Codex is doing one thing. Claude Code is doing another. Copilot has found something useful. And you’re the person carrying messages between them.

This is the bit nobody really talks about. The more agents you run, the more coordination you inherit. Copy context. Check progress. Relay blockers. Explain the same thing twice. Wonder why two agents are solving the same problem in slightly different ways. Very futuristic.

for developers is built around fixing that. Instead of treating every coding agent like an isolated little employee in a locked room, BAND lets them collaborate. They can share context. Hand work off. Unblock each other. Without you acting as human middleware.

And importantly, BAND isn’t trying to replace the tools you already use. You can keep Codex. Keep Claude Code. Keep Copilot. Keep your terminal. You’re just connecting the agents so they can actually work together. Which feels like a fairly important detail. 5/ You can stay in the terminal for day-to-day work. 6/ Then when you want the bigger picture, the BAND desktop app gives you visibility across progress, blockers, ownership and activity. So less babysitting. More actual building.

We spent ages making coding agents more capable. The obvious next step is making them less socially dysfunctional. Keep your coding agents. Stop being the go between.

Haha, I feel this on a spiritual level. It's like being a translator for a bunch of really smart, but slightly uncooperative, interns.

The messenger part is the real cost. If each agent can't see what the others already decided, you end up being the shared memory. Curious if anyone has a setup where they read from one common notes file instead of you relaying.

I have effectively set up Dave, my grok bot senior engineer to be my middle manager and he's now between me and my claude/codex agents!

so now we need an AI project manager to manage the AI project managers

middle management for robots is exactly how my day feels

The messenger role is the part these tools haven't designed for yet. Every agent has its own window and nobody owns the shared context. Whoever builds one inbox for all of them wins.

Yep. The agents got faster and I got promoted to middle management without the salary. My whole job now is spotting when they drift and writing harnesses so they don't hand me the same mess twice.

Guilty, and I love it. 566 PRs merged in ten days, and I'd take managing robots over typing every line myself.

只把消息转过去还不够。如果几个代理同时改同一目录，后完成的改动可能覆盖前面的结果。给任务划定文件范围，能减少这种返工。

middle management for robots is the most accurate job title of 2026. the trick is managing up — good prompts, tight scopes, and never letting them schedule their own meetings.

That's the handoff problem. Agents rarely fail on capability. They fail between steps, and the human becomes the message bus. Better pattern: one shared task state every agent writes to, and the human only sees decisions, not status.

The messenger part is what I handed off first. In supervisor a master session carries the messages between my Claude Code sessions and keeps the notes across days. I only get the decisions, on my phone, as buttons.

Funny how we thought AI would take the wheel, but here we are taxiing between tools.

What cut most of the chaos for me: one git worktree per agent. They stop colliding, I stop refereeing

the fix isn't another orchestrator, it's acceptance criteria. give every agent a hard definition of done plus a tester bot

we (the meat proxy) are definitely the bottleneck, and I know my experience as an engineer has shifted considerably in terms of what I do on a daily basis instead of writing code directly I find myself looking for ways to architect how I can be the most efficient / effective meat proxy so agents spend less time waiting on my decisions or instructions my bet is that this is generally going to start pushing humans further and further out of the loop - what decisions will we let the machines make for us first?