Don't Own the Whole AI Stack. Own the Seams.

On 28 August, OpenAI said it would wind down the contract supplying its models to Cursor, with a proposed shutoff on 12 November, after SpaceX acquired Cursor. Its announcement sets out contractual and terms-of-service reasoning. Read as a commercial dispute, that is somebody else's problem.

It is not only a commercial dispute. It is the latest case in which access to a leading model became entangled with competition between the supplier and the product using it.

The useful reading is neither "pick a different vendor" nor "build your own." You do not need to build the harness. You do need to own the artefacts that make replacing it possible.

Three separate arguments point that way and are worth keeping apart: supply continuity, ownership of what your agents learn, and governance you can evidence. Only the first is about getting cut off.

A Pattern, Not a Grudge

The events are not equivalent and the count matters less than the shape. Anthropic reduced Windsurf's capacity in June 2025, while OpenAI was reported to be in talks to acquire it. It restricted OpenAI's access to Claude that August. In January it cut off xAI's access, which xAI staff had been reaching through Cursor. Now this.

The sequence is structural rather than anybody's misconduct. When the supplier and the buyer of intelligence sell into the same market, access becomes a lever, and levers get pulled. This is hard for every participant, labs included.

The asymmetry belongs in front of a board. For a harness vendor, any one model provider is a share of traffic, and a modest share is survivable. Customers rarely have the same diversification unless they designed for it. Standardise a thousand engineers on one harness, tune the critical workflows against one model pairing, and your exposure is not the vendor's percentage. It is your throughput. You will get ninety days and a blog post.

The Layer Nobody Puts on a Roadmap

A harness is everything around the model that is not the model. The agent loop, the tool definitions, the context assembly, the memory store, the sandbox, the permission checks, the retry and validation logic, the sub-agent boundaries. Cursor is a harness. So is the thing your platform team quietly built last quarter to run nightly reconciliation.

The industry treated this as plumbing for two years, which was defensible while raw capability was the binding constraint. It is less defensible now.

Harness design can materially affect benchmark performance even when the underlying model stays fixed. Terminal-Bench 2.0 explicitly evaluates complete agent systems rather than models in isolation, which makes the harness part of what is being measured.

The implication matters for planning. A meaningful share of what you experience as model quality is decided by code you could be writing.

Just becasue you could doesn't mean you should

Building a harness from scratch is among the worst available uses of senior engineering time. Most of what you would build is undifferentiated: credential vaulting, OAuth lifecycle, multi-user authorisation, audit pipelines, sandbox isolation. None of it distinguishes you. All of it is a security blast radius you now own. Eighteen months later you have a worse version of something you could have bought, three engineers who cannot be reassigned, and a migration you keep deferring.

Nor does open mean safe. An open harness with an ungoverned plugin ecosystem is a software supply chain, and one nobody is staffed to defend is a liability whatever the licence says. Choosing open without funding that work is sovereignty theatre.

…or should you?

The licence fee is the visible cost, and the easy one. Two others sit underneath it.

The first is what you allow the product to accumulate. The valuable exhaust from agentic work is not the prompts. It is the corrections: the record of where your agents went wrong on your systems, what a good outcome looks like in your domain, which tool sequences hold against your schemas. Owning the relevant formats and interfaces lets that record live under your control. Without them, retrieving it is a feature request.

The second is governance. In practice, the recurring obstacles to putting agents in production include evaluation, observability and compliance rather than raw capability. Retrofitting those is expensive and often incomplete, because they only carry weight inside the agent loop. A vendor dashboard tells you what happened. A harness you control determines what may happen in the first place.

Anyone who has built a decision register will recognise the shape. Authority has to be enforced somewhere, and only a layer you can change will hold it.

Own the Seams, Rent the Engine

Harness is not one thing. It is a stack, and ownership belongs layer by layer. The seams are the artefacts that encode your judgement. The engine is the machinery that executes them.

Own the seams:

Context and memory, in formats you can read without the vendor. Documented schemas, stable identifiers, timestamps, provenance and attachments, not only an export button. Anything less should be treated as a portability risk.

Tool contracts, expressed through an open interface such as MCP, contributed as a founding project to the Linux Foundation's Agentic AI Foundation in December 2025. This does not remove migration work: authentication, schemas, error handling and tool-calling semantics still differ. It keeps provider-specific execution details from becoming the contract itself.

The eval suite. Your definition of correct, in your domain, versioned in your repository. The highest-return artefact in the stack, and the one most consistently missing.

The authorisation model and its mapping to agent actions, plus the policy, event schema, retention rules and audit record.

Agent and workflow definitions, as markdown and config rather than proprietary workspace objects.

Routing, so a task can move to the cheapest model clearing the bar on a Tuesday afternoon rather than through a migration project.

Rent the engine: weights and inference, sandbox runtime, the agent loop itself, authentication infrastructure, and the systems that store, search and display your telemetry.

The line is not observability against no observability, or identity against no identity. Own the policy and the record. Rent the machinery that holds them.

Which gives you the test. If the artefacts are portable, replacing the machinery is a project. If they are not, it is a rebuild.

The Half-Day That Turns This Into a Number

Per system, and it fits inside an afternoon.

Export the memory without a vendor SDK or a support ticket, then hand it to a colleague who did not build it. If they cannot interpret it, neither can your next harness, whatever the file extension says.

Count the tools not defined against an open interface. Each is migration work, and it is worth holding the number before somebody asks for an estimate.

Run your three most-used agent workflows on a non-primary model. Degraded output is portable. Crashes mean you built on model-specific parsing behaviour without knowing it.

Search your orchestration code for provider response shapes. Parsing them directly means a model swap fails quietly in production rather than loudly in CI.

Export the governance configuration. If it references vendor workspace identifiers, your controls may not survive a migration, and neither does this year's audit evidence.

Then the question the exercise exists for. If your primary model provider gave you ninety days, what breaks, and how many weeks to recover?

What Appreciates

Models are likely to keep getting cheaper and more substitutable for a growing share of workloads. Anything you build that assumes one specific model is a depreciating asset.

The seams are the other kind. They are where your context lives, your controls are enforced, your corrections accumulate, and a meaningful share of your performance is decided. None of it requires you to build the machinery around them.

So name the person in your organisation who owns the eval suite for your most consequential agent. Not the model choice. Not the vendor relationship. The definition of correct. If nobody owns it, you already know which side of the line you are on. A switch would not be a project. It would be a rebuild, and nobody has priced it.

Originally published on LinkedIn on September 8, 2026.