Open‑Source Coding Agents Are Learning to Run Other Agents

I wanted to see how fast the open-source coding agents are actually moving. So we scored every change that landed on the main branch of nine of them over the last six months, for how much engineering went into it. Two teams do most of the work … in the last 30 days Codex and Qwen Code landed more engineering than the other seven combined. Across the field, the biggest changes are often about having agents run other agents, running them on remote machines, and sandboxing them.

Where the biggest work is going

More and more of the biggest changes let agents run other agents, and their share doubled over the six months. Qwen Code can now hand a subagent's work to an outside agent like Claude Code, Kilo Code added shared boards for agents working on one task, and OpenHands lets an agent launch child conversations.

Many also move the agent onto another machine (remote and hosted runtimes, SSH workspaces, background daemons). Qwen Code, Cline and OpenCode all added SSH support, and Crush added a multiplayer server.

Others limit what the agent can touch, through sandboxing, permissions and trust. That's about a quarter of Codex's biggest work, and many of Gemini CLI's biggest changes are security fixes.

Right now it's Codex and Qwen Code

Codex and Qwen Code landed almost two thirds of the field's engineering in the last 30 days, and they're close to even (Qwen Code edged out Codex over the last 60 days). They got there differently … Qwen Code nearly tripled over the six months, while Codex dipped in early summer and then more than doubled. That's among the open-source agents. Claude Code isn't in the field because its CLI is closed source, and my guess is it would be at the top. Anthropic also defined many of the standards these harnesses build on (MCP, Agent Skills, and Claude Channels, which Crush just added), though ACP itself came from Zed.

Everyone else is well behind. OpenCode lands the most changes after Codex, but they're small ones. Cline, OpenHands and Kilo Code are all growing, and Gemini CLI, goose and Crush slowed down, Gemini CLI the most (we checked it against GitHub to make sure it wasn't a gap in our data).

Harnesses are becoming commodities

My read is that the harness will matter less and the model will matter more. People have strong preferences about harnesses today, but this is the first time I've seen the code moving the other way: the harnesses are adding the plumbing to hand work to each other, and that makes any one of them easier to swap out.

Specializations will probably exist for a while. In our experience Claude is better at frontend and makes nicer UI, and Codex is better at backend and less verbose (at Headline we already pair them, with Claude Code writing and Codex reviewing). How the work gets split across agents is a decision I expect the models to make themselves. The human input is which models you like, plus the spec and the requirements … and over time even the specializations may converge.

Where the agents run is changing too. Right now memory on your own machine is a real constraint (just this weekend I was asked about adding memory to a machine because the agents were slow, and it's the VMs they run that eat it). As the work moves to SSH workspaces and hosted runtimes, the local machine matters less. I'd expect a new kind of hosting provider to show up, one that just runs coding agents, separate from OpenAI and Anthropic. Claude Code or any other harness becomes one way in, and the work runs somewhere else.