Pick the right lane — not everything wants a harness

The right boundary is not "AI or no AI."
It's agentic vs deterministic vs human authority.

A governed operating model uses all three lanes. Picking the wrong lane is how teams overspend on harnesses for work that should have been a script — or hand authority to an agent that should have stayed with a human.

Lane 1 · Agentic

Adaptive, tool-using, exploratory work

When the task needs search, synthesis, iterative inspection, and multi-step judgment against approved tools.

Use it for

  • Discovery, analysis, comparison
  • Drafting and bounded implementation
  • Evidence preparation
  • Guided remediation
  • Runtime diagnosis support
Owned through A governed control loop, bounded tools, visible stop conditions, and the full 5-layer harness.
Trap to avoid "Agent" for work that's actually deterministic. You'll spend more on tokens than the script would have cost.
Lane 2 · Deterministic

Known-path, rule-heavy, repeatable work

When the rules are already known and you want consistency more than open-ended reasoning.

Use it for

  • Code mods and transforms
  • Policy and conformance checks
  • Schema and contract validation
  • Static analysis
  • Repeatable evidence assembly
Owned by Encoded rules, thresholds, predictable traces. No LLM in the loop unless it earns its place.
Trap to avoid Letting "agent" replace this lane just because it's faster to build a prompt. Repeatability and policy posture drift fast.
Lane 3 · Human authority

Approval, exceptions, risk acceptance

Wherever the cost of error or ambiguity is material, the final word stays with a named human.

Use it for

  • Release approval
  • Rollback decisions
  • Policy exceptions
  • Production access changes
  • High-risk customer outcomes
Owned by Named humans with visible authority and accountability. The harness packages evidence; humans decide.
Trap to avoid Collapsing this lane into automation "because the agent looked confident." That's how authority quietly transfers without anyone signing off.
The rule
Match the lane to the work, not the work to the lane you already invested in. A harness only earns its keep when the underlying work is genuinely agentic. Use deterministic pipelines for rule-heavy work and keep authority with named humans for risky or ambiguous decisions.