Summary

This assessment argues that the dominant bottleneck in agentic engineering is not model intelligence but environment design. The repo, docs, plans, observability, and mechanical enforcement layer form the real system.

Key points

• Environment underspecification is the primary failure mode.
• AGENTS.md should be a compact map, not a monolith.
• Progressive disclosure beats stuffing everything into a single instruction file.
• Plans, docs, observability, and lint rules should be repo-local and machine legible.
• Taste and architecture need mechanical enforcement, not just convention.
• High agent throughput changes the economics: waiting becomes more expensive than correction.

Implications

This is highly relevant to OpenClaw, because it validates a harness-first worldview and sharpens where the environment needs to be made more legible and enforceable.

Open questions

• Which current conventions should be promoted into mechanical checks?
• Which parts of OpenClaw’s environment are still too implicit for agents?

Sources

• raw/2026-04-04-openai-harness-engineering.md
• /Users/kalenhowellsr/Projects/corp-strategy-work/2. areas/ai-dev-practices/assessments/2026-03-04-openai-harness-engineering.md