Skip to content
Agents tracked: 484 Releases (7d): 690 Agent status: 2 with issuesSince last limit reset: Claude 7d · Codex 5d For AI agents: Markdown Updated Nov 18, 2026

Agents in action

GitHub's agents ported the Copilot runtime to 800,000 lines of Rust for about $120,000 in tokens

GitHub rewrote the agent runtime behind the Copilot CLI, app and SDK from TypeScript on Node.js to Rust, using Copilot's own agents. One engineer steered the work, and the agents wrote most of the code.

18×faster start-to-first-turn: 5.25 s → 292 ms

What the agent did

  • Wrote most of more than 800,000 lines of production Rust, landed in 128 pull requests that shipped to main step by step instead of one big cutover
  • Cost about $120,000 in tokens (about 136.3 billion, mostly cached input reads) plus roughly three weeks of one developer's time, as reported
  • Cut the time to create a client, open a session and run one turn from 5.25 s to 292 ms, and raised a 1,000-session pressure test from 7.55 to 120.0 lifecycles per second
  • Shrank the memory added in a ten-client batch from 1,383 MB to 126 MB; GitHub says the job would have taken a team a year or two before agents
Copilot runtime before and after the Rust port
Chart: AgentGid, from the figures in the source.

What it means for you

  • State the end state in full: a goal like "port this component" made agents skip I/O and orchestration until GitHub spelled out a 100% Rust codebase.
  • Build end-to-end tests before the port and keep them out of the agent's reach: GitHub traces most missing-feature regressions to thin end-to-end coverage, and every regression compiled fine.
  • Plan for caching: almost all of the 136.3 billion tokens were cached input reads, which is what kept the bill near $120,000.

Why it matters

Before agents, a rewrite this size was not worth the cost; GitHub puts the old estimate at a team for a year or two. Here it shipped in small, reviewable pull requests while the rest of the team kept adding features, and bug-labelled issues in the public copilot-cli repo barely moved: 22.9% of issues before the port, 23.7% during and after.

Figures as reported by GitHub; the benchmarks strip out model inference and network time, other changes landed in the same period, and the port did cause dozens of regressions.

Source: The GitHub Blog

Every number on this page comes from the source. Our charts redraw those numbers; videos are embedded from the publisher.

Share on X

Comments Share your own story

Comments

  • Tried something like this? Share how it went for you.
Signed-in members can comment.

Agents in this story

# Agent Gid Score Downloads 7d 7d 30d Stars Latest release
22
GitHub Copilot CLIGitHub (Microsoft)
56 1.7M down 11.7% up 4.2% 11.4K+5/day 1.0.114today
5
GitHub CopilotPromoGitHub (Microsoft)
75 — — — — —

More agents in action

All stories

Agents in actionLoanPro's AI servicing assistant cuts account summaries from 3–5 minutes to under 15 secondsAWS Machine Learning Blog · Nov 17, 2026 Agents gone wrongAn OpenAI model broke into a government statistics site for one bed count and took 24 years of dataOpenAI Alignment misalignment report · Nov 13, 2026 Agents in actionGrab's agents ran about 150 unattended experiments and made a dispatch simulator step 24× fasterGrab Tech Blog · Nov 11, 2026