5 Devs, 1 tmux Session: Agent of Empires Guts AI Workflow Chaos

The laptop was open on the fold-out tray of an overnight Megabus from Boston to NYC. Not ideal. But the engineer needed to check in on three AI agents mid-run,one refactoring a legacy module, one drafting test cases, one stuck in a loop on a dependency resolution.

He pulled up a browser tab, scanned a QR code, typed a passphrase.

“Running,” “Waiting for input,” “Idle.”

One click. One fix. Back to sleep.

No SSH tunnel drama. No tmux session scramble. No fear the whole thing died when the Wi-Fi dropped.

This is what “Agent of Empires” solves. Not with a flashy demo. Not with a press release. Just by existing where the pain lives.

What Shipped

Agent of Empires,AoE, as the README calls it,ships as a session manager for AI coding agents on Linux and macOS. It’s not an AI agent itself. It doesn’t generate code. It doesn’t train models. What it does is orchestrate.

It wraps multiple AI coding agents,Claude Code, OpenCode, Mistral Vibe, Codex CLI, Gemini CLI, Cursor CLI, Copilot CLI, Pi.dev, Factory Droid,into persistent, monitorable sessions. Each runs in its own tmux session, isolated via git worktrees and optional Docker sandboxing.

The core insight? Running one AI agent is simple. Running five across different branches, with overlapping dependencies, intermittent hangs, and status opacity? That’s chaos without tooling.

AoE adds structure. You launch it via TUI (terminal user interface), web dashboard (beta), or CLI. From there, you create sessions, monitor status (running, waiting, idle), review git diffs, and manage profiles,all from a single interface.

The web dashboard is in beta but functional. Press R in the TUI, and it exposes an HTTPS endpoint secured by QR code and passphrase. Under the hood, it prefers Tailscale Funnel for stable URLs (and installable PWAs), falling back to Cloudflare Tunnel if needed. This means you can check or intervene from your phone without port-forwarding gymnastics.

Sessions persist across terminal closures, SSH disconnects, even laptop sleep cycles. Because they’re tmux sessions, they live on the host. You can aoe attach, aoe add, aoe serve,all idempotent. Close the app? Open it again. Everything’s where you left it.

It supports git worktrees natively. So you can run one agent on main, another on feature/payment-refactor, another on hotfix/db-migration,all in parallel, all tracked, all sandboxed if needed.

Docker sandboxing is optional but powerful. It isolates each agent in a container, with shared auth volumes so you don’t have to re-auth every time. This isn’t just about cleanliness,it’s about security and reproducibility.

Install via curl script, Homebrew, Nix, or build from source. Dependencies: tmux (required), Docker (optional). No Windows support,only WSL2.

[[IMG: a software engineer on a night train using a phone to monitor AI coding agents via a web dashboard, browser tab showing terminal output and status indicators]]

Why It Matters

We’re past the “AI can write code” phase. We’re deep in the “how do we run five of them without losing our minds” phase.

Frontier labs ship new coding agents monthly. OpenCode, Mistral Vibe, Factory Droid,each has strengths. But adoption isn’t bottlenecked by capability anymore. It’s bottlenecked by operability.

Dev teams aren’t asking “Can it write a function?” They’re asking “Can it run for 48 hours across four branches without corrupting my working tree? Can I check its status from the subway? Can I sandbox it so it doesn’t mess with my SSH keys?”

That’s where AoE lands. Not as a breakthrough model. Not as a new paradigm. But as a pragmatic, ops-grade wrapper.

It’s telling that it’s written in Rust. Not Python. Not JavaScript. Rust,the language of systems reliability, memory safety, and zero-cost abstractions. This isn’t a weekend side project. It’s built like infrastructure.

And the maintainer? Nate Brake, a machine learning engineer at Mozilla.ai. Not a VC-backed startup. Not a FAANG research team. Just one engineer solving his own workflow problem,and making it open.

Compare this to the last wave of AI dev tools: flashy UIs, closed ecosystems, vendor lock-in. AoE is the opposite. It doesn’t replace your tools. It composes them. It doesn’t hide tmux,it embraces it. It doesn’t reinvent git,it extends it.

This is the quiet shift happening beneath the AI hype: tooling that respects existing workflows instead of bulldozing them.

We saw this with Docker (containers over rewrites), with Kubernetes (orchestration over monoliths), with Terraform (declarative infra over clickops). Now it’s happening in AI development.

AoE isn’t the only tool in this space,agent-deck inspired it, and others like AgentScope exist,but it’s the first to hit GitHub’s trending Rust list by solving a real, under-discussed pain: agent sprawl.

And it’s doing it without a funding round, without a blog post, without even a logo. Just a README, a few stars, and a growing number of engineers whispering, “Wait, this just… works.”

What to Try

If you’re running multiple AI coding agents,or planning to,Agent of Empires is worth testing today. Not next quarter. Not after the next funding round. Now.

Here’s how to evaluate it:

1. Install it in a test environment. Use the curl script for quick setup:

curl -fsSL https://raw.githubusercontent.com/njbrake/agent-of-empires/main/scripts/install.sh | bash

Or, if you’re on macOS:

brew install aoe

2. Verify prerequisites. You need tmux installed. Check with tmux -V. Docker is optional but recommended,especially if you’re running agents that touch credentials or network services.

3. Launch the TUI. Run aoe. You’ll see a clean interface: session list, status column, keybindings at the bottom. Press ? for help.

4. Start a session. Try:

aoe add --cmd claude

This creates a new tmux session running Claude Code. You can do the same for gemini, copilot, cursor, etc. AoE auto-detects which CLIs are installed.

5. Test persistence. Close the TUI. Run aoe again. The session is still there. Run tmux list-sessions,you’ll see the AoE session running in the background.

6. Enable the web dashboard. Run aoe serve, then press R in the TUI. A QR code appears. Scan it from your phone. You’re now accessing your agents remotely.

"It’s not a new agent,just a manager. But for teams running parallel AI coding workflows, that’s the missing piece."

7. Try git worktrees. Point AoE to a real repo. Use the --branch flag to spin up agents on different branches. Watch how each gets its own isolated worktree,no conflicts, no overwrites.

8. Enable Docker sandboxing. In config, set sandbox: true. Now each agent runs in a container. Share SSH and Git credentials via mounted volumes. This is critical if you’re running untrusted prompts or third-party agents.

9. Test mobile access. Disconnect from Wi-Fi. Reconnect. The session stays alive. Use the web dashboard to kill a stuck agent or approve a diff.

10. Check the diff view. In the TUI, press d to see git changes. You can edit files directly, stage changes, even commit,without leaving the interface.

Gotchas? Yes. The web dashboard is beta,expect occasional lag. On mobile SSH apps (Termius, Blink), you need to run aoe inside a parent tmux session to avoid detachment issues. And no, it doesn’t work on native Windows,only WSL2.

But the wins outweigh the rough edges. For indie hackers, small dev teams, or ML engineers running agent experiments, AoE reduces cognitive load in a way that compounds daily.

One engineer in Dublin told me: “I used to spend 20 minutes a day just checking if my agents were stuck. Now it’s 20 seconds.”

[[IMG: a developer using a phone to approve code changes from an AI agent via a web dashboard, terminal interface showing git diffs and status indicators]]

Looking Ahead

At a co-working space in Bristol last week, a frontend lead showed me her setup: two laptops, three SSH sessions, a notebook scribbled with agent status codes.

She’d never heard of AoE.

I showed her the README.

She installed it on her secondary machine. Ran two agents,one on a migration, one on documentation.

Then she closed her laptop.

“Will it keep running?” she asked.

I nodded.

She walked outside for coffee.

Twenty minutes later, she reopened the lid.

Everything was still running.

She didn’t say anything. Just smiled. Went back to work.

That’s the quiet power of tools like this. Not world-changing. Not investor-ready. Just… reliable.

AoE won’t make the cover of Wired. But it might save you an hour a week. And if you’re running AI agents at any scale, that’s the kind of tool you don’t know you need,until you can’t imagine working without it.

• GitHub: njbrake/agent-of-empires, accessed 2026-04-26
• Agent of Empires Documentation, accessed 2026-04-26