Pocock: Fewer test seams boost agents

The "mock everything" era of software testing is becoming an anti-pattern for AI agents. While humans use seams to manage complexity through isolation, agents are most effective when they can reason about high-integrity vertical slices without the cognitive overhead of brittle internal boundaries.

• –Single-seam problems like building a language or a pure library allow agents to run pure Red-Green-Refactor loops with zero setup friction.
• –Complex apps with many seams (DB, UI, multiple external APIs) force agents to manage state across too many files, increasing the risk of hallucinations and misalignment.
• –This architectural shift favors "Deep Modules" with simple public interfaces and complex internals, which are fundamentally easier for LLMs to reason about than sprawling, shallow abstractions.
• –These principles are being codified in Pocock’s mattpocock/skills repository, which is quickly becoming a standard for disciplined, non-vibe-based agentic engineering.