Invocation Economics (Context vs Cognitive Load)

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

A package contains frequent read lookup and occasional destructive cleanup. Automatic discovery exposes descriptions; user invocation shifts discovery to humans. Host loading differs.

Mechanism

Determine autonomous discovery/composition/manual needs, then configure narrow automatic or supported manual invocation under actual contracts. Document discovery/scope and share references by paths. An index can reduce memory load without assumed automatic calls. Measure real loading/false triggers while tools enforce authority separately.

Bad example

Always trigger everything, or hide cleanup to save context without user discovery; treat loading as permission.

Good example

Make read lookup narrowly discoverable and cleanup manually invoked where supported, documenting entry/effects and checking authority. Record actual loading/frequency/misrouting rather than claiming zero tokens from a flag alone.

Why the change matters

Invocation allocates machine context and human discovery effort. Need-based choices reduce false activation/forgetting; authorization remains distinct.

Observable expectation

Teaching lookup does not activate cleanup; explicitly found cleanup still checks target scope. Shared references have one maintained source and usable links. No load records means no measured token savings.

Limits

Flags/description exposure/calling differ by version. Disabling model invocation is not tool enforcement, and manual calls do not permit arbitrary effects. Frozen zero-context wording is not a measurement here.