I Built an Open-Source Orchestrator for AI Coding Agents

I've been experimenting with autonomous AI coding agents for a while, and eventually ran into a problem:

Running multiple agents is easy. Managing them is a nightmare.

Sure, I could open 10 terminals, start a bunch of coding agents, and hope they don't step on each other's work.

But what happens when they're running on different machines? Who assigns tasks? Who handles failures? And who's stopping an agent from burning through tokens doing absolutely nothing?

So I built SwarmForge.

GitHub: https://github.com/GavinGeizer/swarmforge-oss

What is SwarmForge?

SwarmForge is an open-source orchestration platform that coordinates autonomous coding agents across distributed execution environments.

Instead of managing every agent individually, you give the system a development objective and let it coordinate the work.

The basic architecture:

         SwarmForge
             |
           Manager
             |
      Task Coordination
       /     |     \
    Worker Worker Worker
      |      |      |
     VM     VM     VM
      |      |      |
    Agent  Agent  Agent

Workers operate in isolated environments, and the orchestration layer handles task coordination, worker management, and execution monitoring.

It's also designed to be harness-agnostic, so the infrastructure isn't permanently tied to one AI coding tool or model provider.

Why did I build it?

Hackathons were a big motivation.

I've been participating in university hackathons where we have 24–48 hours to build something substantial.

During these events, I found myself juggling development tasks, AI coding agents, virtual machines, and infrastructure.

I already had access to local servers and remote compute. I just needed a way to make everything work together.

What started as a hackathon tool gradually became a much larger distributed systems project.

The interesting engineering problems

Getting an agent to write code isn't necessarily the hard part.

Making 10 agents work together reliably? That's a different story.

Some of the problems I've been tackling:

Isolation: Keeping agents from interfering with each other's environments.

Lifecycle management: Spawning workers, monitoring activity, and cleaning up idle VMs.

Coordination: Dividing work without creating a mountain of conflicting changes.

Observability: Knowing what the agents are doing without watching 10 terminals.

Cost: Preventing runaway agents from wasting compute and tokens.

And one thing became very clear:

More agents doesn't automatically mean more productivity.

Coordination matters just as much as execution.

It's source available!

SwarmForge is still evolving, but the self-hosted project is available on GitHub.

I'm continuing to improve distributed execution, reliability, supported integrations, and developer onboarding.

If you're experimenting with coding agents, self-hosted infrastructure, or autonomous development, I'd love some feedback.

GitHub:

GavinGeizer / swarmforge-oss

Distributed, self-hosted orchestration for autonomous AI coding agents. Harness-agnostic execution across local and remote workers.

SwarmForge

SwarmForge is a self-hosted orchestrator for parallel AI coding agents running in isolated remote VMs. Your AI lead assigns tasks through the Model Context Protocol (MCP); SwarmForge runs OpenCode workers in Freestyle VMs, tracks their progress, and preserves their results and files.

Website · Setup guide · Releases · Developer onboarding

The coordinator runs on your Linux machine. Each worker gets a separate VM, workspace, and OpenCode session. You supply the VM snapshot, model endpoint, and Git repository. The CLI provides initialization, readiness checks, a terminal dashboard, and artifact retrieval. SwarmForge is source-available under PolyForm Small Business 1.0.0, which restricts permitted business use.

Why distribute coding workers?

A single-machine coding-agent manager runs agent processes against the resources and workspaces of one host. SwarmForge separates the coordinator from worker compute:

• Separate execution environments: each worker runs in its own Freestyle VM with tools from your prepared snapshot. Worker compute…

View on GitHub

Website:

SwarmForge | Multi-Agent AI Coding Orchestration

Orchestrate parallel AI coding agents with SwarmForge. Run isolated OpenCode workers, coordinate tasks through MCP, and preserve results and artifacts.

getswarmforge.tech

I'm especially curious how other developers are managing multiple coding agents. Are you orchestrating them, or are you also drowning in terminal windows?

Built by a computing science student who apparently needed another distributed systems project.

opensource #ai #programming #devtools