Navigation

Preamble

Navigation chooses the route before execution begins: task decomposition, plan shape, model, skill, agent mix, context pack, escalation rule, and stop condition.

---

The First Decision Happens Early

Most agent mistakes begin before the agent starts. The task enters the wrong lane, receives the wrong context, gets the wrong model, or skips the one question that would have changed the work.

Navigation is the routing layer for that first decision. It reads the request, project context, constraints, risk, uncertainty, and capability map, then chooses the smallest workflow that can carry the job while leaving a reason trail behind it.

Routing Can Launder Judgment

The temptation is a clean routing policy that sounds like governance. The policy chooses the lane, the workflow name absorbs the risk, and the operator gets to treat a judgment as a configuration.

Navigation has to expose the reason for a route: direct work, sparring, research, subagent, verifier, halt, or handoff. A route without a reason is automation wearing a uniform.

The Gate Before Execution

Good routing needs friction in the right place. A low-risk edit should move. A high-stakes claim should ground itself. A memory-heavy task should guard against context rot. A philosophy-heavy page should surface its question, assumption, contradiction, test, and punchline before prose begins.

Navigation earns its place when it prevents the expensive mistake. It catches a project that needs sparring before drafting, a branch that needs search before making a claim, a run that needs a verifier before merge, or a shortcut that would let the machine make a judgment I still own.

The First Real Proof

Navigation is proven by fewer wrong-lane runs. The signal is a route corrected before waste accumulates: a simple task kept simple, a risky task escalated, a vague request turned into a sharper question, and a delegated run brought back with enough evidence to trust.