The coordination layer is emerging as the new locus of advantage — rails, harness, operating platforms — while raw capability commoditizes faster than institutions can govern it.

The Tension

The tension is that coordination infrastructure is being shipped as open and neutral while capability access is being closed and credentialed in the same week. Anthropic released Opus 4.7 with identity verification via Persona for certain tiers, including government ID and selfie. OpenAI launched GPT-5.4-Cyber through a Trusted Access for Cyber program with KYC-style verification and $10M in API grants, and The New Stack named the pattern directly. Two frontier labs, on the same day, established that the most capable versions of their models will move through identity-verified channels, not open APIs.

So the barbell is this. Open, interchangeable harness at the base. Credentialed, verified capability at the top. The middle, where traditional SaaS lived, gets squeezed from both sides. If you are building a thin wrapper on a model API, you are in the squeezed middle. If you are building a harness that someone else's agents run inside, or an operating layer where your customers' exceptions become your training signal, you are on one of the two surviving edges.

I think the labs releasing open harnesses and closed capability simultaneously is not a contradiction. It is the same strategy. Give away the coordination layer to harvest distribution. Gate the capability to harvest margin. The trick for a builder is noticing which side you are on, and whether you have chosen it on purpose.

What This Unlocks

For a founder or operator, the practical shift is this. Stop asking which model to use. Start asking what your operation is training, every day, whether you know it or not.

The MIT TR piece gives a concrete mechanic. 50,000 weekly cases times three high-quality decision points yields 150,000 labeled examples per week, generated from normal operations. That is not a training set you buy. It is a training set your workforce produces, if your architecture captures it. Hershey is the operator-level proof. Their decision-intelligence software is projected to deliver $100 million in inventory reduction and $50 million in productivity gain over two years, on top of a $250 million digitization investment. Home Depot acquired Simpl Automation rather than licensing it, because the data generated by warehouse automation is itself strategic. When the feedback loop is the moat, you do not rent it.

The coordination layer unlocks something quieter too. Mastercard integrated with Lobster.cash to let AI agents make card purchases under an agent payment standard. If that standard becomes the default, the agentic economy's spending layer runs through incumbent card rails. Whoever ships the standard first wins the category frame. Builders have a short window to decide whether to run on someone else's rails or carve out their own inside a niche.

Cursor 3 is the same story told at the individual level. The IDE used to be where you wrote code. The new form factor is a process manager for parallel agents. The divided reaction to Cursor 3 is not noise. It is selection pressure. Sequential-editing thinkers hit friction. Parallel-workflow thinkers accelerate. The people who re-architect how they work, not just which tool they use, will pull ahead of the ones who keep typing into a text file.

For a founder today, the decision is not which AI to buy. It is which loop to close. Pick the one part of your operation where a human already adjudicates something a machine executes, and instrument the adjudication. That is where your moat compounds while you sleep.

Watching Next

Three things to watch in the next 90 days, at least one you can check in your own business.

First, whether OpenAI or Anthropic acquires a harness or sandbox company. Modal, E2B, Vercel, and Cloudflare are all candidates. When model differentiation compresses, acquiring the coordination layer is the natural play. If a frontier lab takes a dominant harness position such that more than 60% of production agents run its orchestration, the neutrality of today's rails inverts to platform capture. That is the falsifier for the thesis that coordination is the durable edge.

Second, the Qwen 3.6-35B-A3B benchmark results on SWE-bench Verified or equivalent. If open-weights reaches parity with Opus 4.7 and GPT-5.4-Cyber on production agentic workloads, credentialed access becomes the only real moat, and that moat is fragile to regulatory change. The shape of the barbell depends on this number.

Third, in your own business, count how many decisions your team made this week that a software system executed downstream. Now ask how many of those decisions were captured as labeled examples. The delta is your operating-layer gap. You can measure it today. If that number is zero, you are renting intelligence, not building it.

Underweighting

The coordination-layer thesis has a structural weakness I have not adequately addressed. Every vendor building coordination infrastructure is doing so as a distribution play, not a product play. OpenAI's Agents SDK is model-agnostic because OpenAI's models do not need coordination-layer moats. They need distribution surface. Cloudflare's Code Mode MCP reduces token costs because Cloudflare profits from edge compute and storage, not from MCP abstraction. The rails are being built by entities whose actual value accrues elsewhere. That means the neutral coordination layer is structurally a gift to the market, and the companies positioning themselves as 'coordination layer businesses' are competing to own something the most powerful players are giving away.

The Jensen Huang argument on Dwarkesh I borrowed cuts the other way. Nvidia has bilateral supply-chain lock with $250 billion in committed purchase orders. Temporal and Modal have no comparable lock. Kubernetes is the honest analogue, and Kubernetes became infrastructure nobody makes margin on. If coordination middleware follows that trajectory, the right read is not 'own the coordination layer' but 'the coordination layer will be free by 2027, and value will re-concentrate at model capability and application-layer proprietary data.' For a founder, that implies the opposite of this essay's prescription. The feedback-loop thesis is correct. The loop should be built at the application and data layer, not at the coordination layer itself.

I should also be more direct about what the Hershey case proves and does not. Hershey's decision intelligence system is a notification layer for supply-chain operators, real-time alerts on performance data. That is not the 150,000-labeled-examples-per-week mechanic the MIT TR piece describes. The $100 million and $50 million figures are two-year projections from a vendor and have not been validated by independent audit. I stacked these as operator-level proof, but they are operator-level aspiration. A reader who follows the link to Supply Chain Dive will notice the gap between my framing and the source. I will owe a correction in six months if the numbers slip.

Bottom Line

Capability is becoming cheap and gated at the same time. The defensible middle is the loop where your operation converts daily decisions into proprietary signal, wrapped in rails you either own or have deliberately chosen to run on. Today, pick one workflow where a human already adjudicates something a machine executes, and write down what it would take to capture that adjudication as training data. If you cannot answer in a paragraph, you do not have an operating layer yet, and the clock on building one has started.