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.