GitHub Rewrites the Copilot Agent Runtime Into 800,000 Lines of Rust, Built Largely by Its Own Coding Agents
GitHub says it rewrote the Copilot agent runtime into more than 800,000 lines of Rust using its own Copilot coding agents, at a roughly $120,000 token cost.
Overview
GitHub has rewritten the agent runtime that powers GitHub Copilot from TypeScript into more than 800,000 lines of production Rust, according to GitHub. In a September 16 blog post, Stephen Toub, a Distinguished Engineer at Microsoft, wrote that “AI agents wrote most of the code,” with the rewrite carried out using the GitHub Copilot app and the Copilot CLI across 128 pull requests that shipped incrementally rather than in a single cutover, according to GitHub.
What We Know
The Copilot agent runtime originally ran on TypeScript, Node.js, and the V8 JavaScript engine, and it backs the GitHub Copilot CLI, the GitHub Copilot app, and the GitHub Copilot SDK, according to GitHub.
Toub wrote that the project “would have taken a whole team of developers a year or two before agents” but was “completed primarily by a single developer, in only a few months, all while the rest of the team continued to greatly expand the runtime’s capabilities and reach,” according to GitHub.
The porting effort ran for roughly fourteen-and-a-half weeks, according to GitHub. By August 21, the runtime reached 100% production Rust, totaling 832,378 lines of production Rust plus 468,689 lines of Rust unit tests and 174,675 lines of end-to-end TypeScript tests, according to GitHub. By September 14, the team had traced and fixed “dozens of known port regressions,” most of them correctness bugs, according to GitHub.
Toub put the project’s token spend at roughly 136.3 billion tokens, at a cost of about $120,000, plus what he estimated as roughly three weeks of his own time, according to GitHub.
On a benchmark of 1,000 one-turn session lifecycles, Toub reported the pre-port TypeScript CLI completed 7.55 lifecycles per second, compared with 57.45 for the Rust runtime running out-of-process and 120.0 running in-process, according to GitHub. Measuring memory added during a ten-client batch, he reported the pre-port process tree peaked at 1,383 MB above baseline, versus 247 MB for Rust out-of-process and 126 MB in-process, according to GitHub.
Analysis
GitHub’s disclosure adds to a short list of large, agent-assisted Rust migrations that companies have detailed this year. Bun, the JavaScript runtime that Anthropic acquired last December, announced in July that it had ported more than 535,496 lines of Zig to Rust, with founder Jarred Sumner using a pre-release Claude model to do much of the work, as reported by The New Stack. OpenAI has separately disclosed that two engineers using Codex and GPT-5.5 rewrote Habitat, a storage service underpinning products including ChatGPT, from Python into Rust during the second quarter of 2026, with the new service “handling 95% of our production requests” while using six times less CPU and fifteen times less memory than its Python predecessor, as reported by The New Stack.
The New Stack noted that GitHub’s approach differed from Bun’s: Bun pushed a large, mostly parallel port before merging the rewritten runtime back into its main codebase, while GitHub shipped the Copilot runtime piece by piece across those 128 pull requests over roughly fourteen-and-a-half weeks, continuing other development on the product throughout, as reported by The New Stack. The outlet also pointed out that both GitHub and Anthropic are vendors of the coding agents they credit with making the rewrites possible, calling their accounts “inherently self-interested,” as reported by The New Stack.
What We Don’t Know
Toub cautioned that the benchmarks are workload-specific, noting that the Rust runtime is not universally “15.9x faster,” the ratio between the 7.55 and 120.0 lifecycles-per-second figures from the 1,000-lifecycle test, according to GitHub. GitHub has not published further technical detail on OpenAI’s Habitat rewrite, which The New Stack reported based on OpenAI’s own disclosure rather than an OpenAI primary-source document, as reported by The New Stack.