The same Claude model, but inside Cursor it can locate bugs, edit code, and run tests; swap it into Claude Code or Codex and it suddenly seems "dumber," spitting out long guesses. This is a counter-intuitive observation circulating in the developer community.
What this is
A widely-discussed article on Juejin proposed a framework: Agent ≈ Model + Harness.
Think of an AI Agent as a person: the Model is the brain — responsible for understanding, reasoning, judgment, and code generation. The Harness (literally "horse harness," here meaning the entire engineering system wrapped around the model) is the body, senses, and nervous system — responsible for reading files, calling tools, organizing context, and managing the execution flow.
The Harness contains at least three things: context engineering ("which code should the model see"), tool orchestration (which tool to use at what moment), and the execution loop (observe → think → act → feedback). Same tools, different orchestration — wildly different outcomes.
Industry view
Mainstream read: AI competition is sinking from the "model layer" into the "system layer." DeepSeek has begun publicly disclosing its Harness design in Code Agent benchmarks, and now exposes interfaces to plug its own models into Claude Code, Pi, and other Coding Agents. That is an implicit admission that the model alone no longer separates winners from runners-up.
The dissent still matters. Engineers in the comments push back: the absolute capability gap between models is real. Pair Claude Sonnet with the same Harness as a small open-source model, and the code quality still sits in a different league. The Harness only amplifies existing intelligence — it cannot conjure it from nothing.
Another risk: the engineering dividend of the Harness gets flattened by open-source frameworks. Once mainstream Coding Agents publish their core patterns, the cost of imitation collapses, and competition snaps back to the pace of raw model progress.
Impact on regular people
For enterprise IT: when procuring AI tools, "whose model sits underneath" is no longer the only question to ask. What needs pressure-testing is how context is managed, how tools are orchestrated, and how errors feed back. These "invisible parts" are what actually determine outcomes.
For individual professionals: swapping tools with the same model underneath can swing the experience by several multiples. Learning to tune the tool — how to feed context, how to decompose tasks, how to verify results — may be worth more than memorizing any single tool's feature list.
For the consumer market: future AI product differentiation will increasingly hide in the small details — a prompt template, a feedback loop, a flow design. On the surface, they all look similar; the difference lives in the invisible system.