11. Harness Engineer
Evidence label: Capability. Harness engineering names a production discipline; the evidence reviewed does not support a stable standalone Harness Engineer job family.
What they do
Build the infrastructure that keeps AI agents "under harness," under control. As agentic AI systems generate code, take actions, and operate with increasing autonomy, harness engineers build the systems that ensure they stay within architectural constraints, produce coherent output, and don't accumulate entropy over time.
The three pillars
1. Context engineering (knowledge infrastructure) Not one-off prompts, but a continuously updated knowledge base embedded in the codebase. Agents know your conventions, architecture decisions, and domain context. Dynamic access to observability data and documentation.
2. Architectural constraints (agent guardrails)
- LLM-based watchdog agents that review generated code
- Custom deterministic linters enforcing your specific architectural patterns
- Structural tests (ArchUnit-style) that run automatically
- Pre-commit hooks that reject code violating established constraints
3. Entropy management (drift prevention) Periodic agents that scan the codebase for: outdated documentation, architectural violations that slipped through, abandoned patterns that reappeared, inconsistencies introduced by multiple agents working in parallel.
The core insight
Without a harness, AI agents produce code that individually looks fine but collectively drifts away from your architecture, your patterns, and your documentation. The harness is what makes "AI generates most of the code" sustainable at scale rather than a path to unmaintainable systems.
Organizational home
Platform engineers, staff engineers, developer-productivity teams, and architecture groups currently own this work. Treat Harness Engineer as a search keyword and capability profile, not as a forecast that every organization will create the title.
Required skills
Software architecture, linter/static analysis tooling, LLM orchestration, observability, codebase knowledge management, entropy detection patterns.