An 11-hour unattended agent run: 2 of 6 ideas shipped

Claim

The first measured unattended run in the corpus ends in human review, not auto-merge. Repo conventions went into .agents/skills and the agent was asked for one unmerged pull request with numbers. This supports [[2026-10-skills-agents-md-convergence]] (Google uses the .agents/skills layout) and qualifies [[2026-10-codex-software-factory-grokbot-google]], where the Google entry was only a launch announcement.

Evidence

All spans from the Google Developers Blog post, 2026-10-13 (Weale, Yang, Biernat):

• "Argon ran unattended for 11 hours overnight"
• Prompt: "Open one pull request against main that the ADK maintainers would approve, with before and after numbers in its description." and "Don't merge it."
• "6 proposals led to 2 shipped changes after code review." The team "rejected the other four changes."
• "That division of labor is why the workflow worked: the agent explores, the maintainer subtracts."
• "Session preparation drops by about 40% in long sessions and package import time drops by 21%."
• The 158-turn run goes "from 65.6 seconds to 56.1 seconds".

The post's caveats (the 40% scope, a synthetic author-written benchmark, test breakage from lazy-loading) were reported by a reader from a summary and were not checked against the page, so they are not claimed here.

Open questions

• Whether /goal mode in Antigravity matches the goal modes in Codex or Claude Code.
• Reproducibility outside Google.

Sources

• https://developers.googleblog.com/optimizing-a-200000-line-python-codebase-overnight-with-gemini-4-argon/ — blog, Mon Oct 12

AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.