Harness Engineering vs Prompt Engineering
Writing good prompts and building good execution environments are different things
In 2023-2024, how well you craft prompts was the key to AI utilization. In 2025-2026, that's shifting.
Limits of Prompt Engineering
No matter how good your prompt is, if the model can't read files, it can't do code review. If it can't run tests, it can't verify correctness. If it can't see error messages and retry, it has to get it right on the first try.
Prompts are "instructions." Harness is "capability." Perfect instructions mean nothing without capability.
What Harness Engineering Determines
Tool access — Can the model access file systems, browsers, APIs, databases?
Feedback loops — Can it see execution results and retry? Do error messages reach the model?
Context management — During long tasks, what information is kept vs discarded?
Guardrails — Are there mechanisms to preemptively block dangerous actions?
Orchestration — In what order and under what conditions are multiple model calls executed?
What SWE-bench Showed
The same model scoring 2-3x differently depending on the harness is well-documented on SWE-bench. Devin, Claude Code, Cursor, SWE-agent all use the same base models but perform differently. The difference is the harness.
This isn't saying prompt engineering is useless. Prompts are one component of the harness — not the whole thing.
How It Works
Prompt Engineering — optimizing instructions to the model (system prompt, few-shot, chain-of-thought)
Harness Engineering — designing the entire execution environment (tools, feedback loops, context, guardrails)
Prompts are one layer of the harness — prompts alone cannot read files, run tests, or retry on errors
SWE-bench proves it — same model, different harness, 2-3x performance difference