Thin vs Thick Harness

Model-lab voices argue capability makes scaffolding redundant while harness builders show large harness-only gains, and the resolution is that harness thickness should track model capability and how verifiable the task is.

Evidence: weak

Anyone building an agent has to decide how much machinery to put around the model. Too little and the agent loses track, loops or declares success early; too much and the harness burns context, encodes outdated workarounds and becomes its own maintenance burden. The field has not settled where the line sits.

The thin side. Latent Space reports Boris Cherny and Cat Wu (Claude Code) and Noam Brown arguing that the value lives in the model and the harness should stay thin. It also reports that METR and Scale's SWE-Atlas found harness differences within noise. Anthropic's own guidance is that components encode model weaknesses and should be removed as models improve, and a claude.dev post confirms a cut of more than 80% of Claude Code's system prompt with no measurable loss (see harness assumptions go stale). A secondary history adds that WorkOS got better results after deleting about 95% of its skill text, which is unverified.

The thick side. Jerry Liu is reported as saying the harness is everything, and Latent Space cites harness-only gains across 15 models. LangChain moved an agent from 52.8% to 66.5% on Terminal-Bench 2.0 by changing only the harness, and preprints report ranking reversals of up to 38 points from harness choice alone (see harness as a performance lever). swyx concludes that harness engineering is increasingly real.

The evidence is weak on both sides. The Cherny, Brown and Liu positions come from a page that showed them truncated, with no primary source fetched, and the 95% deletion anecdote is secondary.

A workable resolution is that the right thickness depends on how capable the model is and how verifiable the task is. Where success can be checked mechanically, a thin harness with strong checks can work; where it cannot, unattended loops degrade, as described in unverifiable loops rot. For agents that take operational actions, the task is often partly verifiable (a health check either passes or not) but the blast radius is large, which argues for thick sensors and thin, auditable action paths: rich read-only diagnostics plus a deliberately narrow action verb set. That is where this debate meets the minimal general tool surface argument.

Sources

  • Latent Space (AINews), 5 Mar 2026 — https://www.latent.space/p/ainews-is-harness-engineering-real
  • Viv Trivedy, LangChain, 17 Feb 2026 — https://www.langchain.com/blog/improving-deep-agents-with-harness-engineering
  • Polomodov, 1 Oct 2026 (secondary history) — https://polomodov.tech/en/2026-10-01-agent-harness-evolution
  • Thariq Shihipar, claude.dev, 24 Jul 2026 — https://claude.dev/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models/