Why Coding Agents Improve More From Harness Design Than Models
LangChain pushed a coding agent to Terminal-Bench's top 5 without changing the model — the gains came entirely from harness design.
A widely cited data point: LangChain moved a coding agent from rank 30 to the top 5 on Terminal-Bench without swapping the underlying model. The entire gain came from the scaffolding wrapped around it — what Anthropic's own agent team calls the "harness," defined as everything that turns raw model capability into completed work: system prompt, tools, memory, loop, checks, sandbox.
The field is framed as moving through three eras: prompt engineering (wording the ask well), context engineering (curating what fits the context window and avoiding "context rot," where attention thins as more tokens pile in), and now harness engineering — designing the loop, tools, roles, and gates the model operates inside.
Four practical building blocks stand out: keep the base loop minimal and grounded in real feedback before adding complexity; manage context with Reduce/Offload/Isolate patterns to keep the window high-signal; separate generation from evaluation so an agent doesn't grade its own work and declare victory prematurely; and decouple the Brain (model plus loop), Hands (ephemeral execution sandbox), and Session (durable state) so each piece can be upgraded or swapped on its own schedule — which is also what makes cost-based model routing feasible.
The takeaway for engineers: the model is a rented, commoditized component that vendors can change quarterly without notice. Durable engineering leverage lives in the harness you own, not in the API you consume.