The Thesis
Knowledge + structured execution > modelAgent reliability improves more from giving agents the right knowledge and constraining their execution than from switching to a larger model.
What This Means
Knowledge
Not “more data” — curated, structured domain knowledge delivered to the agent at the right time:- Which testing patterns work for this framework
- What dependencies are available and how to use them
- What the project conventions are
- What common failure modes look like
Structured Execution
Not “better prompts” — infrastructure that shapes agent behavior:- Deterministic preprocessing before the LLM acts
- Tool configuration that guides tool selection
- Execution loops with built-in checkpoints
- Judge feedback that catches failures early
> Model
This doesn’t mean models don’t matter. It means that for a given model, you get more reliability improvement from knowledge and execution infrastructure than from upgrading to the next model tier.Agents Are Workflows
Before the thesis can be tested, the object it applies to has to be named honestly. An agent is not a magic black box — it is a workflow. Each step is either deterministic (build, lint, test, measure coverage) or an AI step (reason about an error, generate code, decide what to fix next). What people casually call “an agent” is usually just the AI step — one node in a larger pipeline.Evidence
Code Coverage v1
The first experiment showed two independent axes of improvement — knowledge injection and prompt hardening — both of which operate on infrastructure, not model choice. The PetClinic “model floor” (92-94% coverage regardless of variant) demonstrates that the model’s prior knowledge creates a ceiling that only infrastructure can differentiate.SkillsBench (External)
SkillsBench (Feb 2026) found that 2-3 curated skills improve agent performance by +16.2 percentage points on average. Comprehensive skills actually decrease performance by -2.9pp. This validates “curated > comprehensive” — structure matters.Stripe Convergence
Stripe arrived at the same structure independently, at a scale that makes it hard to dismiss. Their Minions system merges over a thousand pull requests a week containing no human-written code, against hundreds of millions of lines of Ruby on a platform handling more than $1 trillion in annual payment volume. They call the pattern blueprints — deterministic steps interleaved with agent reasoning steps, which is the diagram above under a different name. As Stripe’s Alistair Gray puts it: “Blueprints combine the determinism of workflows with agents’ flexibility in dealing with the unknown.” The detail that matters most for this thesis is why they needed it: their stack is uncommon and their libraries are homegrown, so the model does not know their codebase. Rules files supply the knowledge and blueprints supply the structure — knowledge and execution structure compensating for what the model does not have. Two independent groups converging on interleaved determinism is better evidence than either alone. Sources: Minions: Stripe’s one-shot, end-to-end coding agents (Alistair Gray, February 2026), and Anup Jadhav’s The walls matter more than the model.The Equation
Naming History
This concept has gone through several names:How to Apply
If you’re building agent systems:- Start with knowledge — What does your agent need to know that it doesn’t?
- Structure the delivery — Skills > flat files > nothing
- Add execution constraints — Deterministic preprocessing, judge feedback loops
- Then consider the model — Upgrade only after infrastructure is solid