Coupled in training, decoupled in deployment: our call on whether the AI harness can be split from the model

Key takeaways

The question is whether the harness can be pulled apart from the model underneath it. The harness is the control layer that wraps the model. It is the part users actually deal with in Claude Code or Codex. This is one of the most valuable unsettled questions in AI right now.

One camp says the two separate cleanly. That would turn models into interchangeable parts, priced like a commodity. The other camp says the two have to be tuned together. In that view, whoever holds the point of integration takes the profit.

This piece puts the evidence from both camps on the same table for the first time. Neither camp wins. Our call is a dividing line that moves: the harness and the model are coupled in training and decoupled in deployment.

The commoditization of models started with easy tasks and is climbing up the difficulty scale. At the hard end, the premium for a natively integrated model and harness is still getting thicker. The evidence for the hard end is Anthropic's revenue curve. The evidence pool for this piece holds no clean band-by-band measurement of how much quality drops after a model swap.

Where this debate sits on this publication's map:

The fight: the same model, in a different shell, can be a different product

The debate starts from an engineering fact that runs against intuition. How well an agent performs depends on the shell as well as on how smart the model is.

The same Claude 4 Sonnet does far better inside Anthropic's own Claude Code than inside Cursor, and better again in Cursor than in GitHub Copilot. The researcher Nathan Lambert likened using that last combination to riding with a half-drunk driver (Interconnects / Substack). Academic work named the effect back in 2024: agents are a new kind of end user, so interface design is itself a performance variable. On the SWE-bench of that time, the same model behind a different interface scored several times higher (arXiv).

This May the effect got a controlled measurement. Harness-Bench is a diagnostic test that holds the task, the budget and the evaluation protocol fixed and swaps only the harness. Run across several model backends, it produced 5,194 execution traces, and the same set of models scored up to 23.8 points apart overall depending on the harness (arXiv). That is one self-reported paper, not yet replicated. Its direction matches a July paper: changing only the harness, with the model held fixed, more than doubled an open model's score on a computer-use benchmark (arXiv).

If the shell matters that much, the next question is whether shell and core are bound together. Two camps give opposite answers, and neither is made up of minor players.

The case for separation: models are parts, and value sits in the harness and the data

What is being claimed. The separation camp's chain of reasoning runs like this. The harness is open source and not tied to any model, so switching models means changing one line of configuration. The user's memory, workflows and context all live in the harness, not with the model maker. So any squeeze a model maker applies, whether a block or a price increase, does not destroy demand; it pushes customers to a cheaper supplier. Value and bargaining power migrate out of the model layer into the harness, routing and data layers.

Who's betting on the claim. Peter Steinberger wrote OpenClaw, an open-source harness that works with any model and has more than 180,000 stars on GitHub (Lex Fridman Podcast). Databricks' Matei Zaharia launched Omnigent, a common layer that maps Claude Code, Codex and Cursor onto a single interface (X / @matei_zaharia). On the enterprise-software side, Databricks CEO Ali Ghodsi argues that the model layer ends up like TSMC: valuable, but interchangeable (BG2 Pod / Altimeter). The heaviest backer is Microsoft CEO Satya Nadella. In a May interview Nadella explicitly rejected the idea that model and harness are inseparable. He said Microsoft runs one multi-model harness across GitHub and Microsoft 365, with models hot-swappable, and that automatic routing is one of the company's largest continual-learning efforts.

How hard the evidence is. This year the separation camp picked up evidence at three levels.

At the behavioral level, three companies or users acted on the claim. After Anthropic blocked OpenClaw users from running through Claude subscriptions, one user showed publicly that a mix of open models could rebuild a US$200-a-month workflow for about US$15 a month (TechCrunch). Perplexity ran internal evaluations and then made a model costing roughly half as much the orchestrator in its product. Perplexity's CEO summed up the shift as "the real product is now the harness around it" (Latent Space / AINews). The legal AI company Harvey dropped its in-house legal model because frontier models beat that model on Harvey's own evaluations. Harvey now sells the layer that orchestrates any frontier model into workflows for law firms, and in March it raised money at a US$11 billion valuation (CNBC).

At the academic level, one paper shows that an automated search over harness designs can beat hand tuning, and that the configurations it finds carry over to other models (arXiv).

At the vendor level, Microsoft confirmed to Axios that it is putting a fine-tuned DeepSeek V4 into Copilot Cowork as a low-cost option, expected within weeks (Axios).

Where the claim is weak. The universal version, that models are a commodity, runs into one official number. Anthropic put US$30 billion of annualized revenue into its own fundraising press release and described growth of roughly 80 times in two years (Anthropic). If models were truly interchangeable, a single lab should not keep capturing so much of the revenue growth. The commoditization camp itself had named this counter-evidence as a signal to watch. A second limit comes from security. Steinberger noted that resistance to prompt injection rises with model strength, so for sensitive tasks being able to switch models is not the same as being willing to switch (Lex Fridman Podcast). One more limit needs stating plainly. The separation camp's academic evidence is all single self-reported papers, and its behavioral cases mostly come from one user or one company.

What would prove this wrong. The separation camp named its own falsifier. If Omnigent-style common layers do not become a de facto standard within a year, and startups building harnesses that sit across models keep failing at a high rate, the argument that models are becoming parts can no longer be treated as a leading indicator.

The case against separation: the integrator takes the profit, and bolt-on stacks can't catch up with native ones

That is one half of the table. The other half starts from an old theory of how industries organize.

What is being claimed. The integration camp's chain of reasoning comes from Clayton Christensen. While a product is not yet good enough, integrated players win; modular players take over only once the product overshoots what customers need. Agentic workflows are far from good enough today, so the harness has to be tuned together with the model to work at all. The dream of sitting on top and plugging in any model does not work yet. Profit in the value chain flows away from modular links toward integrated ones, which means toward model makers that build their own harnesses.

Who's betting on the claim. Analyst Ben Thompson argued in March that the Microsoft-style dream, sitting on top while the models stay interchangeable, cannot be realized; at least for agentic workflows, model and harness have to be integrated for now (Sharp Tech / Stratechery). OpenAI CEO Sam Altman said in an interview that OpenAI no longer treats the harness and the model as two separable things, and that tool calling is moving from a bolt-on to a capability built in during training. Ethan Mollick, who writes the One Useful Thing newsletter, added a product fact: the three most complete harnesses today, Claude Code, Codex and Antigravity, all come from the model makers themselves (Ethan Mollick).

How hard the evidence is. The integration camp's evidence also comes at several levels.

At the revenue level, Anthropic officially reports US$30 billion in annualized revenue. Beyond that total, one product line shows a model maker running its own harness: Claude Code. Anthropic's official February figure for Claude Code was US$2.5 billion annualized, with enterprise subscriptions up fourfold since the start of the year. Claude Code's downloads also ran more than 20 times those of its rival Codex (Interconnects / Substack).

At the behavioral level, Microsoft has the most reason of anyone to go modular, yet Copilot Cowork runs on Microsoft's own cloud and is built on Claude's harness design (Sharp Tech / Stratechery). Companies building products close to the model have felt the opposite pressure. Cognition has had to rebuild Devin, its coding agent, for each new model, and with some model generations some of Devin's capabilities went backwards (Stratechery).

At the mechanism level, the 23.8-point Harness-Bench gap reads the other way for this camp: the pairing effect between model and harness is real. A model cannot be dropped into just any harness and work as well.

In hindsight, a 2023 prediction held up: integration would be the biggest roadblock, and the products that survived would be the model makers' own first-party ones. Three years later, both parts of that prediction had come true (Interconnects).

Where the claim is weak. The integration camp's most concrete prediction has not come true in four months. In March Thompson said that layered companies like Harvey could be wiped out by a model maker with a markdown file and a bit of reinforcement learning. Harvey then raised money at a US$11 billion valuation, and its annualized revenue climbed from US$195 million to about US$300 million in five months. The layered company did not die; it changed how it makes a living.

Microsoft is also loosening. On June 16, when Nadella announced Copilot Cowork's general availability worldwide, multi-model support was in the first sentence (X / @satyanadella). The same day Axios confirmed the timeline for the DeepSeek low-cost tier. Microsoft 365 Copilot already lets users choose Claude, and more than 1,500 customers use models from both OpenAI and Anthropic on Azure (Stratechery).

Our read is that the old gap between revenue and profit also applies. Capturing revenue growth is not the same as capturing the margin. High growth has also come with heavily subsidized competition.

What would prove this wrong. The integration camp's own falsifier: an enterprise-grade harness that is not tied to any model appears, and it performs no worse. That would overturn the claim that the two cannot be separated.

Both camps on one table: the dividing line is not a date but a frontier that climbs

Four months ago the two camps talked past each other. Stack the new evidence on top and a shape appears: each camp is describing a different part of the same problem, and one variable divides their territory: the difficulty band of the task.

Start with where coupling happens. Altman's statement that OpenAI no longer treats harness and model as separable points to the mechanism. A model is tuned with reinforcement learning inside its maker's own harness. So the experience premium of the native pairing is real. That explains why the same model behaves like a different product in Claude Code and in Copilot. It also explains why Cognition rebuilds its product for every model generation. The coupling happens in training.

Now look at where decoupling happens. Perplexity changing its orchestrator model, Harvey's model selector, Microsoft's low-cost tier and automatic routing, the OpenClaw user's US$15 rebuild: all of these happen at the deployment layer. By that point the harness has already broken the task down, routed it and wrapped it in guardrails. Which model runs a given step then becomes a question of price and performance. The academic result that harness design can be searched automatically and carried across models shows that the craft of the harness layer is not a moat owned by the model makers (arXiv). The decoupling happens in deployment.

Both of these hold at once, and together they rewrite Christensen's line that modularity arrives only once a product is good enough. The original theory treats that moment as a point in the future, and Thompson's version puts it a long way off. The evidence points instead to a frontier that moves by the difficulty band of the task. For low-band tasks, such as writing a weekly report or sorting email, models have already become parts: buyers compare prices rationally and open models can do the work. For high-band tasks, such as closing a hard ticket or changing a large codebase, the premium for native integration is still getting thicker, and Anthropic's revenue curve is the price evidence for that stretch. The live question about commoditization is how far up the difficulty bands it has climbed.

Which camp is right therefore depends on where a given task sits. How fast the frontier climbs has a measurable deciding variable: a public comparison, band by band, of how much quality drops after a switch to a cheaper model. That comparison table is empty today. The evidence pool behind this piece contains no clean enterprise-scale measurement that directly answers how many points performance drops when the model changes under the same harness. That table is what this debate most needs. Its absence is why both camps can only trade anecdotes.

Our call

There is no single end state for whether the harness can be decoupled from the model. Because harness and model are coupled in training, the native model-and-harness pairing keeps an integration premium on hard tasks, and the profit stays with the model makers for now. Because they are decoupled in deployment, the commoditization of models keeps climbing from easy tasks upward, and the profit flows to the routing and orchestration layer. This call has three falsifiers that can be checked within twelve months. ① If a verifiable enterprise case appears where hard tasks move to a non-native model with no loss of quality, the integration-premium call is dead. ② If, after Microsoft's low-cost tier launches, highly complex tasks do not migrate to it and Omnigent-style common layers do not become a standard, the call that commoditization is climbing gets downgraded. ③ If multi-model routing visibly spreads out Anthropic's revenue concentration, the leg of the call that says the integrator takes the margin breaks.

What this changes, and for whom

• AI investment allocation: the slogan that models will commoditize so the application layer is the place to be, and the slogan that model makers win and middleware dies young, are both universal claims, and this year's evidence has trimmed both. The version that holds sorts by difficulty band: the case for natively integrated products sits at the high bands, and the case for the routing and orchestration layer sits at the low bands. The three falsifiers above serve as the gauge for how fast the dividing line climbs.
• Teams building agents: coupling in training means the native pairing's lead on hard tasks is not marketing; on those tasks, swapping in a cheaper model carries a real quality risk. Decoupling in deployment rewards architectures where the choice of model for each step is a routing decision, since each band that commoditization climbs takes another slice out of cost. The one internal instrument that matters is a table of quality loss from model swaps across a team's own task mix. The evidence pool behind this piece holds no published version of that table.
• Microsoft: Microsoft is the live experiment in this debate. In March it was the textbook counterexample, the company most eager to go modular that had tied itself to a single harness. In June it announced multi-model support and brought DeepSeek into a low-cost tier. Two readings matter after the low-cost tier launches: the share of tasks routed to it, and whether highly complex tasks move with them. The first measures how fast the frontier is climbing. If the second happens, the claim that harness and model cannot be separated has really been broken.
• Model makers' strategy: Anthropic closed subscription access for third-party harnesses and got, in return, a user publicly demonstrating a US$15 rebuild. Our read is that restrictions do not keep low-band demand. They speed its move to open-weight models. Model makers' pricing power lives in the high bands. Fighting over the low bands as a revenue defense amounts to teaching customers how to migrate to rivals.

To close, look at the tree once more. This piece updates the Harness value jump branch. The path that says the two cannot be split narrows to coupling in training: its premium is real, but it lives in the high difficulty bands. The path that says they can be split gets this year's first anchors at the behavioral and measurement levels, but its territory so far reaches only the low and middle bands. The fork stays alive. It has changed from a choice between two end states into a measurement question: how far up the frontier has climbed.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.