Claude Code team: ~90% of code written by Claude, verification loops unlock massive parallel agents
ClaudeDevs · x · 2026-11-17
The Claude Code team (Addy Osmani) explains how they equip coding agents to verify their own work with evidence:
- Roughly 90% of the team's code is written and reviewed by Claude. Verification loops are the key to delegation: without them you can only run one or two Claudes at a time; with them, dozens or hundreds in parallel.
- In internal tests on structured file generation, adding a separate grader checking work against a rubric improved task success by up to 10 points, with the biggest gains on the hardest problems.
- Core insight: verification that used to be expensive—tests, benchmarks, specs, real end-to-end runs—is now cheap to request from an agent, and once generated these run as software repeatedly at little extra token cost.
- Loops can feed back into prevention: repeated failures (e.g., a check on logs leaking user data) can be fixed with a hook or an explicit rule in CLAUDE.md.
The post also covers end-to-end, property and differential testing, and how the team keeps its verifiers honest.
More from coding & agent
- After a weekend with Fable 5.2, going back to an Opus 5.5-only work setup is jarring — nathanborror · 2026-11-17
- Sharing AI Workflows Should Mean More Than Sharing the Prompt — Far-flungOutdoors · 2026-11-17
- Google ships A2UI 1.0, first stable version of its framework-agnostic generative UI protocol — rseroter · 2026-11-17
- Platform Engineering Mentions Surge ~150% as AI Coding Tools Force Governance — craigsdennis · 2026-11-17
- Everything Antigravity: open-source agent OS with MCP security harness and 40-70% token savings — ProfessionalBridge89 · 2026-11-17
- Stanford publishes free 6-lecture CS329Z video series on building AI agents — stanfordnlp · 2026-11-17