← OrreryThe Ecliptic is an Orrery publication

Cover Story · Parameter desk

After the Human Leaves the Gate

GitHub now lets a machine's approval count toward merge rules, a month after agents started more than 40% of its pull requests. Firms that merge with fewer people report fewer reverts; outside studies of agent code find more corrective work. The two records count different things, because a revert needs a person to notice.

110 million
pull requests started from an agent's first draft on GitHub in September 2026, more than 40% of the month's total, against 17% in March
1.62×
the odds that a merged agent pull request drew a verified follow-up fix within 30 days, against human pull requests in the same repositories
69.6%
of those follow-up fixes to agent code were written by the same agent; 27.4% by a person
> 40%
fall in the Claude Code team's revert share from June to September 2026 while merged pull requests more than doubled (self-reported)

The gate

Approval became optional on the largest code host

On 28 October 2026 GitHub made Copilot code review approvals generally available. When an administrator enables them, Copilot's approval counts toward a repository's required approvals.

  • Off by default: the approval is dismissed when new commits are pushed, and administrators can restrict it to chosen file paths.
  • Volume: agent-started pull requests rose from 17% of the total in March to more than 40% in September, about six times March's volume.
  • Merge rates: 90% of agent-involved pull requests merged within two weeks in September, against 83% in February and about 86% for pull requests with no detected agent.
  • A different rule inside one firm: the Claude Code team ties approval to a risk tier set by a classifier and holds that an agent's approval never counts as a person's.

After the merge

Outside studies find more corrective work, not more work overall

0.511.522.533.5Hazard ratio, agent lines vs human lines (1 = no difference)All modifications1.11Corrective maintenance1.49Bug fixes1.51Security fixes1.20Reverts1.42
Post-merge outcomes of agent-written lines relative to human-written lines, Cox hazard ratios with 95% confidence intervals; teal marks intervals that exclude 1. Source: Xia and Miller, arXiv:2607.09902, 182 repositories, May 2025 to May 2026, as reported in Parameter PAR-0027. Not illustrative.

Across 182 repositories from May 2025 to May 2026, agent-written lines were no more likely than human lines to be modified at all, but the modifications were more often fixes.

  • Corrective maintenance: hazard ratio 1.49; bug fixes: 1.51. Both confidence intervals exclude 1.
  • Reverts: 1.42, with an interval from 0.57 to 3.54, so not distinguishable from human code.
  • Review link: each 10-point rise in a repository's share of unreviewed merges was associated with about 6% more agentic maintenance burden.
  • These are 2025-vintage agents, one or two model generations behind those in use now.

Why the records disagree

A revert needs a person to notice

0%4%8%12%16%Pull requests reverted within 90 days of mergeHuman baseline11.5%OpenAI Codex6.1%5.4% of its pull requests human-reviewedDevin14.5%
Share of pull requests reverted within 90 days of merge, December 2024 to July 2025, 37,623 pull requests in 2,807 repositories. Copilot, Cursor and Claude Code were not significantly different from the human baseline and are not shown. Source: Kraishan, arXiv:2609.17598, as reported in Parameter PAR-0027. Not illustrative.

Revert share, the figure the firms report, is produced by human attention: someone must spot a defect, trace it to a change and choose to back it out. When defects are fixed forward instead, they never become reverts.

  • In the one dataset with both review coverage and reverts by agent, the agent with the least human review, OpenAI Codex at 5.4% of its pull requests, also had the lowest revert rate.
  • Our reading (Parameter): fewer reverts can mean fewer defects, or fewer people looking. Revert share alone cannot tell them apart.
  • The metric to track instead: corrective-maintenance rate and the share of defects fixed forward, split by review regime.
The Ecliptic, Issue No. 017, cover analysis, drawn from Parameter PAR-0027 (November 4, 2026). Platform figures are GitHub's own; firm figures are self-reported; study figures are as published by their authors. Interpretation is Parameter's.
THE ECLIPTICan Orrery publication
Content is produced by AI systems with human editorial oversight.