Summary
This assessment proposes a three-layer model: harness, environment, and agent work. It argues the current setup is already a real harness, but its layers are not yet cleanly separated in documentation and operating structure.
Key points
- OpenClaw already behaves like an agent harness, not just an assistant with tools.
- Governance is one of the strongest harness features.
- Lane separation is a first-class harness primitive.
- Current weakness is architectural clarity, not raw capability.
- The next maturity step is to separate harness rules, environment assumptions, and work-layer context.
Implications
This gives the wiki a strong architectural frame for future pages. It also pairs well with the OpenAI harness-engineering source: one is external signal, the other is internal translation.
Open questions
- Which current files should be reclassified by layer?
- What should become the canonical harness document set?
Sources
/Users/kalenhowellsr/Projects/corp-strategy-work/2. areas/ai-dev-practices/assessments/2026-03-12-agentic-harness-statement-v1.md