AI Coding Wins with Verification, Harnesses, and Structure
Shift AI coding from fast generation to rapid verification using harnesses with sensors; structure functions to reveal intent; reject 'software brain' by prioritizing precise data definitions over total AI legibility.
Verify Fast to Outpace AI Generation
In agentic programming, speed comes from verifying multiple approaches quickly, not generating one slowly. Generate five options and check all in an afternoon to beat teams waiting weeks for feedback. Invest in review surfaces like tests, type checkers, and automated gates over perfect prompts—'verified' now means automated checks plus human judgment where needed. Use tools like Claude Code or Codex CLI with inner harnesses for reliable agent throughput. Train AIs by example so diffs approve themselves first time; senior engineers shape harnesses and teach teams, turning oversight into compounding leadership.
Fundamentals endure: keep changes small, build guardrails, document ruthlessly, verify every change. Distinguish 'vibe coding' (ignore output) from agentic engineering (structured verification). Programmers train AIs to write proper software, passing skills to juniors.
Harness Engineering Adds Objective Sensors
Computational sensors like static analysis and tests in harnesses provide deterministic assurance where LLMs handle fuzzy rules. Agents fix every static warning exhaustively, unlike humans who slack. Birgitta Böckeler's experiments show sensors boost reliability; formal checks outperform LLM inference for objective rules.
Structure Functions to Convey Intent Explicitly
AI models infer from tokens, names, and local context—not deep semantics—so explicit structure matters. Avoid line-count rules; extract code fragments into functions named for their 'what' (intent), hiding 'how' (implementation). A four-line wrapper returns a new program concept, localizing change and clarifying logic. Wrong boundaries cascade comprehension issues; good ones build vocabulary for humans and AIs alike.
Reject Software Brain for Precise Definitions
' Software brain' sees the world as controllable databases, reducing people to data points—fueling AI backlash over surveillance. Like lawyers crafting protocols, programmers define behaviors, but law's non-determinism (litigation) contrasts code's rigidity. Internal data fails AI due to inconsistent definitions; precise, tended conceptual models are essential for LLMs and genies. Making yourself 'legible' to AI (dumping notes/emails) aids querying but risks caricaturing self. Writing refines your brain—AI drafting cripples it; understanding grows in your mind, not transient sessions.