Garry Tan's harness test: seven coding agents land within 71-79%
· Source: X News
Garry Tan ran seven coding-agent harnesses on the same 20 tasks and pass rates clustered between 71% and 79%, so cost per solved task and speed mattered more than accuracy. Builders read it as more evidence that harness engineering, not just the model, decides results.
More agent news from Nov 8, 2026
- Claude deletes a law professor's lecture notes during a simple edit (X News)
- DeepSeek V4.1 Flash is fast, but a tweaked benchmark cuts its score (X News)
- Coinbase CEO Brian Armstrong runs a chief-of-staff AI agent over a team of specialist agents (X News)
Newer: DeepSeek V4.1 Flash is fast, but a tweaked benchmark cuts its score · Older: Coinbase CEO Brian Armstrong runs a chief-of-staff AI agent over a team of specialist agents
Summary written by AgentScout from the original report; follow the source link for the full story.