Shaping the flow of ideas at the frontier of cognitive systems.
—
The same AI model scored 78% on a coding benchmark in one environment and 42% in another. The model was identical—Claude Opus 4.5 in both cases. The difference was the harness: the software infrastructure that determines what an AI agent can perceive, remember, and act on. One harness gave the model access to a persistent file system and conversation history. The other reset its memory after every task. The 36-point performance gap proves what most teams building with AI still underestimate: infrastructure determines outcomes more than model capability.
A harness is the scaffolding around an AI model—the execution environment, the way it accesses tools, how it manages state between tasks, whether it remembers past interactions. Think of it as the difference between giving someone a problem to solve in a quiet room with reference materials versus shouting instructions through a door and hoping they remember what you said three hours ago. Same person, radically different performance. The harness architecture that surrounds a model shapes its effective intelligence more than the parameters inside it.
—
Five architectural decisions compound into lock-in. First: execution philosophy. Does the AI run on your local machine where you and it share the same file system, or does it run in the cloud where it observes your work through API calls? Claude Code chose local-collaborative—human and agent share a sandbox, enabling real-time feedback. Systems built on OpenAI’s Codex infrastructure chose cloud-autonomous—the agent works in isolation, trading human coupling for reliability at scale. Second: state management. Does the agent maintain context across sessions by reading git history (a record of all changes to code over time) and progress files, or does it start fresh every time? Persistent state creates momentum. Ephemeral state avoids coupling. Third: context preservation. Does conversation history thread continuity across days, allowing the agent to learn from your corrections, or does each task get a clean slate? Fourth: tool integration. Can the agent read your file system directly and run shell commands (instructions executed in a text-based computer interface), or must every action go through formal API contracts (predefined interfaces that enforce strict rules about how software components communicate)? Fifth: multi-agent coordination. When you need multiple AI agents working together, are they peers collaborating in a shared workspace, or workers in a hierarchy where one orchestrator dispatches tasks to specialized subordinates?
None of these choices is obviously wrong. Cloud execution enables running thousands of agents in parallel without asking humans to provision machines. Stateless design prevents cascading failures when one agent’s corrupted memory poisons the next task. API contracts enforce architectural clarity. Hierarchical coordination scales to complex workflows. But each choice pulls the entire system toward a coherent philosophy, and the philosophies don’t mix. You can’t bolt stateful memory onto a stateless cloud architecture without breaking the assumptions that made the cloud architecture reliable. You can’t add tight human-in-the-loop feedback to a system designed for autonomous operation without sacrificing the autonomy that justified the complexity.
—
The lock-in arrives slowly, then all at once. In year one, you choose an execution philosophy because you need to ship something. By year two, your workflows depend on the state management that philosophy enabled—developers rely on the agent remembering yesterday’s feedback instead of re-explaining context every morning. By year three, your team optimizes for the strengths of your chosen architecture. Developer ergonomics improve. Iteration velocity climbs. By year five, your organizational muscle memory has formed around the harness. Switching would mean retraining every workflow, rewriting every integration, re-learning which tasks the system handles well. The compounding is silent and nearly irreversible.
Cursor, the AI coding assistant, reached $2 billion in annual recurring revenue with 100% of that income flowing through API calls to cloud model providers. That valuation isn’t just a bet on current revenue—it’s a bet on the cost of customers switching away. Five years of workflow dependencies, tooling assumptions, and team habits built around a cloud-autonomous philosophy creates gravitational lock-in. The technical choice becomes an economic moat. The infrastructure doesn’t just run the AI—it captures the customer.
—
The performance gap between harnesses isn’t a bug to be fixed by the next model release. It’s a structural property of how systems accumulate intelligence over time. One architecture learns from your corrections and builds context across sessions. Another guarantees reproducibility by forgetting everything between tasks. Both strategies produce value, but they produce different kinds of value, and the difference compounds.
Technical debt isn’t just badly written code. It’s the weight of accumulated architectural assumptions—the invisible cost of every infrastructure choice you made when moving fast. The harness you choose today determines which conversations your AI can remember, which tools it can access fluently, and how much friction sits between human intent and machine action. You’re not selecting a vendor. You’re selecting your future constraints.
—
Source: “Same model, 78% vs. 42%. The harness made the difference.” — Nate’s Newsletter, March 6, 2026.
Text summarized and optimized using Anthropic’s models and reviewed by a human.