Pulse
2026-09-18-v3.0
Dataset
2026-10-27-catalog-ai-v2
Ruleset
b2e8889c66f7
Scanned
2026-10-27
Window
February–September 15, 2026
Catalog tracking observation. This repository is measured with the same fixed window and ruleset, but it is not part of the published benchmark or its aggregate claims.

Textbook on reinforcement learning from human feedback

01Result

natolambert/rlhf-book ↗

natolambert/rlhf-book AI coding statistics: 64.6%

of 246 commits in February–September 15, 2026 were committed by an AI agent — a floor, not an estimate
64.6% agent · top contributor 52.8%14 contributors
team

A small group is using agents. Adoption has spread beyond one person but is not project-wide.

159
agent commits
87
other considered commits
14
agent contributors
52.8%
largest contributor

02Drivers

What shapes this result

Mode, timing, and detected signatures show whether the result reflects broad adoption or one concentrated workflow.

Assisted or autonomous

159assisted
0autonomous

Assisted — a developer’s identity authored the commit and the agent added itself as co-author. Autonomous — a bot identity authored the commit without a human identity attached.

Agent share by month

Agent-attributed commit share by month
MonthAgent shareConsidered commits
2026-01100%1
2026-0280.6%36
2026-0371.4%42
2026-0475%48
2026-0526.2%42
2026-0688.9%27
2026-0776.2%21
2026-0852.2%23
2026-090%6

Peak month: 100%. Bars are scaled to that peak, not to 100%.

Which agents

AgentVendorCommitsShare
Claude CodeAnthropic15864.2%
GitHub CopilotGitHub10.4%

03Evidence

Inspect the matches

Public commit samples let you inspect the signatures behind the aggregate result.

Matched commit samples

  • 10421433 on GitHubBonus 1: What's With the Shoggoths? (#518)Claude Code
  • 636f102f on GitHubCI: cache slide decks, reuse print formats on HTML-only changes, paralleliClaude Code
  • c808e0c5 on GitHubAdd "Web Version vs. Physical Book (Errata Fixes)" section to introductionClaude Code
  • 54adf05d on GitHubFix uninitialized loss in ORM and PRM forward pass (#402)GitHub Copilot

04Method

What entered the denominator

The result begins with commit history, removes merges and routine automation, and reports literal agent signatures as a lower bound.

Exclusion ledger

Commits in window246
Merge commits removed−0
CI / dependency bots removed−0
Considered246

Bots are excluded from the denominator entirely. Agent attribution requires a literal supported signature; no code-style inference is used.

Continue exploring

Claude Code commitsBrowse repositories with detected Claude Code signatures →GitHub Copilot commitsBrowse repositories with detected GitHub Copilot signatures →Python AI coding statisticsCompare evidence across Python repositories →GitHub AI commit analyzerInspect another repository with the same evidence model →
Reproduce this observation

Run the same detector locally.

Deterministic signature matching; no code-style inference. Squash merges can remove trailers, so this result remains a measured floor.

npx qmmit-cli natolambert/rlhf-book

Scan another repository

Try