FIND-20260331-014 — Harness Engineering: Why the Best AI Engineers in 2026 Stopped Writing Code

Nav Toor (@heynavtoor) shared a thread on harness engineering — the emerging discipline of designing the infrastructure, constraints, memory, and feedback loops that wrap around AI coding agents rather than writing code directly. The core insight is that the model is commodity; the harness is moat. Key components include CLAUDE.md/AGENTS.md context files, MCP servers, skills (progressive context disclosure), sub-agents as context firewalls, lifecycle hooks, and back-pressure mechanisms (tests, typechecks). Evidence: LangChain improved coding agent accuracy from 52.8% to 66.5% by changing only the harness, not the model. OpenAI built a 1M-line app with zero human-written lines. This is not speculative — ODS already practices harness engineering through its ADLC pipeline architecture.

ODS is already a practitioner of harness engineering: CLAUDE.md files per service, skills system (~/.claude/skills/), agent memory (agent-memory/), sub-agent delegation pattern, and feedback loops via lessons-learned.md. This thread validates the ODS architecture and provides a vocabulary for what the team is building. Actionable improvements: (1) Audit CLAUDE.md files for conciseness — recommendation is under 60 lines; current files are longer and may be degrading agent performance. (2) Review skills for progressive disclosure — skills should not all load at once. (3) Add lifecycle hooks (pre-commit typechecks, build verification) to tighten the back-pressure loop. (4) Consider the 'never write code without a failing harness signal' principle as an extension of TDD. Directly affects: all ADLC agents, dev agent, security agent, ba agent.

Harness Engineering: Why the Best AI Engineers in 2026 Stopped Writing Code

Source

ODS Impact

Security Review

Tags