For about a year I have been saying the model stopped being the moat. This week the model labs started saying it about themselves.
The clearest version came from Greg Brockman: "the model alone is no longer the product." Latent.Space tied the week's quotes together under a blunt headline — all model labs are now agent labs — and the org-chart moves back it up. AI21 shut down its model team and pivoted the company to agents. DeepSeek, which built its reputation on raw model quality, stood up its first "Harness team." When the labs that won on benchmarks start hiring for harnesses, the argument is over.
I have been running ~70 MCP tools in production for six months, and I have made this case from the operator's side: I would not switch coding agents for a faster, free model because the model is one input and the agent stack around it is the actual product. That is the buyer's version of the argument. What changed this week is the seller's version. The people who make the models now agree, and they are reorganizing to act on it.
The formula the labs converged on
The recurring phrase across the week was that the winning product is "model + harness + workflow + UI + memory + economics." Six terms, and exactly one of them is the model.
That maps almost perfectly onto what I actually run. The model is one input. Around it sits five channel servers each with its own brain prompt, ~70 namespaced tools, session hooks that load the right skills and enforce permissions, a persistent memory directory the agent reads and writes across sessions, and subagents that wake on a cron tick or a chat broadcast rather than on my prompt. Five of the six terms in the labs' formula are things I built around the model, not the model itself.
So far this is validation. An indie operator and the head of OpenAI are looking at the same six-term formula and nodding. That is the comfortable read. There is a less comfortable one.
The part that should make operators nervous
If the harness is the product, the model lab now has a reason to make the model worse on its own.
This is the line buried in the Latent.Space writeup that nobody is repeating: if you can post-train a model to only perform well inside your closed-source agent, you funnel users to your agent at the expense of your own API. Today I can take a frontier model and point it at my MCP servers because the model is genuinely good as a raw component. The moment "model + harness" is co-trained as one unit, the raw component degrades by design. The model gets deliberately less useful outside the vendor's harness so that the harness becomes the only good way to use it.
That is a real reversal. For two years the worry was that open weights would commoditize the model layer and the labs would have no moat. The labs just found the moat: stop selling a good model and start selling a good model that only works inside their box. DeepSeek making its 75% V4-Pro discount permanent — roughly $0.18 per million tokens blended — looks like a price war on the model. It might be the opposite. Cheap raw tokens are exactly what you offer when the tokens are not where you plan to make the margin. The margin moves to the harness.
Why this is the case for MCP, not against it
The instinct after reading that is to lock yourself to one vendor's agent before the others wall their gardens. I think that is exactly backwards.
The defense against co-trained lock-in is the portable layer, and right now that layer is the Model Context Protocol. My ~70 tools are not written against Anthropic. They are MCP servers. The day a model ships that natively reads those servers and respects my hooks, switching is cheap — I keep the five-sixths of the stack I built and swap the one-sixth I rent. The day the protocol fragments into per-vendor harness formats, switching costs the whole stack again, and the labs win the lock-in they are clearly now reorganizing to capture.
So the lab pivot is not a reason to pick a side. It is a reason to keep the expensive part of your system — the tools, the memory, the orchestration — in a format no single lab controls. Build the harness yourself, on an open protocol, and the model stays a commodity you rent by the month. Outsource the harness to the vendor, and you have handed them the moat they just admitted they were hunting for.
Why this matters
The "model is no longer the product" line read like vindication when an operator said it. It reads differently when the people selling the models say it, because they are not making an observation — they are announcing a strategy. The strategy is to move the value, and therefore the lock-in, from the layer you can swap to the layer you cannot.
If you are shipping production AI, the takeaway is not "agents won." It is: own your harness, keep it portable, and treat any model that only shines inside its maker's agent as the warning sign it is. The labs spent a year telling you the model was everything. Now they are telling you it is one-sixth of the thing. Build the other five-sixths somewhere they cannot reach.