9 AI Coders, 1 Rust Cage: Agent of Empires Stops Branch Burns

The night-shift engineer in a Vancouver co-working space didn’t say “session management” when I asked why he’d installed yet another tool last week. He said, “I was losing tabs.”

Three terminals. Two tmux panes. A phone propped on a coffee cup running a remote SSH client. One AI agent rewriting a migration script. Another stuck in a loop on a feature branch. A third silently corrupting the main worktree while he slept.

He showed me the command he ran at 2 a.m.: aoe add --cmd claude. One line. No setup. The agent spun up in a clean git worktree. He went back to sleep. Woke up to a diff, not a disaster.

That’s the pitch for Agent of Empires,a Rust-based session manager for AI coding agents that just landed at the top of GitHub’s trending list in the Rust category. It doesn’t train models. Doesn’t generate better code. Just keeps your agents from eating your codebase.

The Deployment

Agent of Empires (AoE) is a terminal-first tool that lets developers run, monitor, and manage multiple AI coding agents,Claude Code, OpenCode, Mistral Vibe, Codex CLI, Gemini CLI, Copilot CLI, Pi.dev, Factory Droid, and Cursor CLI,all from a single interface. It uses tmux under the hood, so sessions persist even when the terminal disconnects or the app closes. Each agent runs in its own isolated environment, optionally sandboxed with Docker, and tied to a specific git worktree.

Installation is minimal: curl a script, use Homebrew, or build from source. Once running, the TUI (text-based user interface) shows all active sessions, their status (running, waiting for input, idle), and lets you jump in with a keystroke. Press R, and a web dashboard launches,accessible from any browser, installable as a PWA, protected by QR and passphrase.

The web interface isn’t just a status board. It renders the actual agent terminal in the browser. You can type into it, switch sessions, review diffs,all from a phone. The key shortcut Ctrl+b d detaches from a session without killing it. Reopen aoe, and everything’s where you left it.

The tool doesn’t replace your agents. It wraps them. You still need Claude Code or Copilot CLI installed locally. AoE detects what’s available and gives you a control plane. No orchestration. No workflow automation. Just visibility and isolation.

It’s not for teams pushing code through CI/CD pipelines. It’s for individual engineers,often working solo or in small pods,who run AI agents locally and need them to behave.

[[IMG: a developer in a Berlin apartment using a phone to check the status of AI coding agents running on a laptop, terminal window glowing in the dim light]]

Why It Matters

The quiet story here isn’t about AI coding. It’s about terminal culture making a comeback.

Three years ago, the big push was “no-code AI.” Then “AI IDEs.” Now? Engineers are going back to the terminal,not because they hate GUIs, but because they need precision.

Agent of Empires doesn’t try to make AI agents smarter. It assumes they’re flawed. That they’ll hang. That they’ll ask for input and wait forever. That they’ll edit the wrong file in the wrong branch.

So instead of building a better agent, Nate Brake,the creator, and a machine learning engineer at Mozilla.ai,built a better cage.

This is the post-hype phase of AI tooling. The phase where we stop pretending agents are reliable and start building around their instability.

Think of it like early Docker. People didn’t adopt containers because they made apps faster. They adopted them because they stopped apps from breaking each other.

Same logic here. Git worktrees prevent branch collisions. Docker sandboxes prevent dependency clashes. The TUI dashboard prevents attention fragmentation.

And the mobile access? That’s not a gimmick. I’ve watched engineers in Dublin, Melbourne, and Austin use SSH clients on their phones to check agent status during commutes. They’re not coding on the train. They’re triaging.

One told me: “If my agent’s stuck waiting for input, I can reply in 30 seconds and buy myself four more hours of runtime.”

The web dashboard being in beta tells you where this is headed. Right now, it’s a convenience. Once it stabilizes, it becomes a productivity multiplier,especially for devs who work across devices.

But the real signal is in the stack: Rust, tmux, git, Docker. This isn’t a Silicon Valley startup tool. It’s a *nix native tool built by someone who lives in the terminal. No telemetry. No cloud dependency. No vendor lock-in.

That’s why it’s trending. Not because it’s flashy. Because it’s done.

What Other Businesses Can Learn

If you’re running AI agents locally,especially across multiple projects or clients,Agent of Empires solves a real ops problem. But it’s not plug-and-play. Here’s how to evaluate it:

First, your team needs to be in the terminal already. If you’re using hosted IDEs or browser-based agents, skip this. This tool assumes you’re comfortable with tmux, git worktrees, and Docker. If you’re not, the learning curve will outweigh the gains.

Second, assess your agent sprawl. Are you or your engineers running more than one AI coding tool? Are they working across branches or repos simultaneously? If yes, session drift is probably costing you time. One dev in a Bristol fintech startup told me he was spending 15 minutes a day just reconnecting to hung sessions. After AoE, it was under two.

Third, consider isolation. The tool’s Docker sandboxing is optional, but you should enable it. Agents make mistakes. They install wrong packages. They modify system files. Running them in containers limits blast radius. One indie hacker in Toronto showed me how he uses AoE to test agent-generated deployment scripts in a sandbox before letting them near production.

Fourth, mobile access works,but only if you set up Tailscale Funnel or Cloudflare Tunnel. Tailscale is smoother: it gives you a stable URL, and your PWA stays working across restarts. Cloudflare Tunnel is the fallback. Either way, you need an account and basic networking setup. Don’t expect mobile access out of the box.

"The tool doesn’t improve agent output, it prevents agent sprawl."

Fifth, Windows users: you’re out unless you use WSL2. The README is clear,native Windows isn’t supported. The tool relies on POSIX process handling and tmux, which don’t translate. If your team is Mac or Linux, no issue. If not, this won’t help.

Finally, monitor the web dashboard’s progress. Right now, it’s beta. But once it stabilizes, it becomes a remote-work enabler. Imagine spinning up an agent on your laptop, then checking its progress from your phone while walking the dog. That’s the vision.

For engineering leads, the lesson is broader: as AI agents become routine, the tools that win aren’t the ones that make agents smarter, but the ones that make them safer to run. Think linters, not compilers. Seatbelts, not engines.

[[IMG: a software engineering lead in a Toronto co-working space demonstrating the Agent of Empires web dashboard on a tablet, with a terminal window and documentation open nearby]]

Looking Ahead

The last thing the creator said before I logged off was: “I just wanted to stop losing my sessions.”

No vision statement. No roadmap for “AI orchestration.” Just a personal pain turned into a public tool.

I’ve seen that before,in the early days of Git, of tmux, of Homebrew. Tools built for one person that end up serving thousands.

Agent of Empires won’t replace your IDE. Won’t write your next feature. But if you’re running AI agents locally, it might just keep you from rewriting your main branch at 3 a.m.

And that, for now, is enough.

• Agent of Empires GitHub Repository, accessed 2026-04-26
• Tailscale Funnel Documentation, accessed 2026-04-26
• Cloudflare Tunnel Setup Guide, accessed 2026-04-26