Architecture AI Updates: October 26, 2026

1. A Long-Running Coding Agent Pattern: Scoped Skill Files, a Persistent Goal, and a Hard No-Merge Rule

InfoQ. InfoQ covered Google’s unattended 11-hour run of Gemini 4 Argon against the Agent Development Kit (ADK) Python repository: over 200k lines across about 800 files, with 16,892 unit tests. The harness has a few reusable parts. Repository conventions went into .agents/skills files, and the agent loaded only the skill relevant to its current task so the context window would not fill up. Antigravity’s /goal mode kept the objective intact across context compaction. The prompt required an unchanged public API and model-facing content, a regression test with every change, and a single PR with before/after numbers that the agent was not allowed to merge. Maintainers accepted two optimizations: filtering events before copying session history (about 40% less session-preparation overhead in long sessions) and lazy imports (21% faster cold import). Together they cut a 158-turn tool-calling run from 65.6 s to 56.1 s. Reviewers rejected four other proposals for reliability, maintainability, compatibility, or too little measured benefit, which shows human review is still where agent output gets limited. Source