Three cards from two operators. Two themes: the bifurcating AI application stack and agent governance constraints.
Decision model economics, the cost and accuracy case for specialized classifiers
Tom Tunguz published "AI Comes for the If Statement" on September 21, 2026. The same day, Simon Willison reviewed TypeSafe's Jev model from a different angle. Neither piece references the other. Both are examining the same shift: the move from frontier text-generating models to narrower models that output typed results.
Once an AI system is hardened and repeats at scale, specialized decision models replace frontier models at a fraction of the cost and higher accuracy documents the economic mechanism. Tunguz argues the AI application economy is bifurcating: frontier models for discovery and architecture, specialized decision models for production classification tasks that recur at scale. His evidence is specific. TypeSafe's Jev tested on 98 hand-verified production email threads achieved 80% accuracy, compared to 47% for the production generative LLM classifier. Cost reduction: 76 to 209 times compared to frontier model pricing. The mechanism is architectural. Decision models skip autoregressive decoding, evaluate candidate answer logits directly in a single attention pass, and return typed confidence scores. The task shape (binary decision, classification) matches the computation required, removing the token-generation overhead that dominates frontier inference cost.
"We use the largest frontier models to train new models... But once a system is engineered & hardened, running it thousands or millions of times benefits from narrower AI"
Decision models that output typed numbers instead of text remove the reasoning trace, making decisions fast and cheap but unexplainable captures the governance cost that the economic case omits. Willison's point: text-generating LLMs produce a readable reasoning trace by default. Decision models don't. Jev accepts text in and returns a floating point number out. You can verify the decision; you cannot inspect the path to it. For any context where audit trails or decision explainability are required, that's a precondition failure, not just a missing feature.
"Jev doesn't even give you that: put in all the text you want, the only thing you're going to get back is a floating point number."
The two cards read together form the full tradeoff map for teams considering specialized decision models: large cost and accuracy gains in the right conditions, at the cost of interpretability in contexts where that matters.
- Tom Tunguz, Once an AI system is hardened and repeats at scale, specialized decision models replace frontier models at a fraction of the cost and higher accuracy. Once an AI system is hardened for a repeating classification task, specialized decision models replace frontier models at 76 to 209 times lower cost and near-double accuracy.
- Simon Willison, Decision models that output typed numbers instead of text remove the reasoning trace, making decisions fast and cheap but unexplainable. Decision models that output typed floats instead of text sacrifice the reasoning trace, making decisions fast and cheap but unexplainable.
Agent governance, MCP as the four-capability floor for constrained deployments
Willison published a short commentary on September 20, 2026, responding to MCP skepticism on Hacker News. The argument is structural. Terminal agents with full shell access have unrestricted trust; they don't need a governance protocol because no constraint is imposed. The moment an organization needs to constrain an agent, four specific requirements emerge: which external services it can call, how it authenticates without touching API keys directly, how users connect and authorize further services, and how every action is logged. MCP satisfies all four through a standard interface rather than custom per-deployment implementation.
MCP's value is governance for constrained agent deployments, not capability for autonomous ones names this clearly. The protocol's purpose is not capability. It is the governance layer that makes controlled agent deployments possible in enterprise, regulated, or multi-tenant environments. YOLO terminal agents skip it because they have no governance floor to satisfy.
"If you want to operate something that's less YOLO than that, you'll find yourself wanting [these four capabilities]"
This is distinct from MCP 2.0's stateless HTTP transport reduces server implementation to a single request-response cycle, making agent tools easier to build, scale, and audit than shell access, which documents MCP 2.0's transport-layer redesign. That card covers the protocol's architectural efficiency. This card covers its organizational purpose.
- Simon Willison, MCP's value is governance for constrained agent deployments, not capability for autonomous ones. MCP's value is governance for constrained agent deployments: service access control, auth abstraction, user connection UI, and audit logging that fully autonomous agents don't need but enterprise deployments do.