Generic harnesses still help on unseen org-specific cases, but not as a long-term substitute

ruslansv · x · 2026-07-27

The author adds that elaborate generic harnesses are still useful when dealing with organization-specific distributions the labs never saw, and for quickly adapting between model releases.

But they argue such scaffolding is not a durable substitute for models that have already absorbed long-horizon agent behavior. The mistake, they say, is the same kind of error as believing better prompt engineering would permanently outrun better base models.

Related event: General Agent Harnesses Fall Short Against Long-Horizon RL Models(2 posts)→

Original post →

More from coding & agent

coding & agent channel →