The model proposes; the runtime decides
A capability gateway keeps tool use governed even when model behavior remains probabilistic.
Tool calling gives a model reach. Reach is useful, but it is not authority.
A model can suggest that the system send a message, update a record, or retrieve private context. Whether that action is allowed should be decided by deterministic code with access to identity, permissions, budgets, and current product state.
One boundary for every capability
Instead of giving each tool its own scattered safety checks, route requests through a capability gateway:
decision = gateway.authorize(
actor=actor,
capability="profile.write",
arguments=arguments,
budget=run_budget,
)
if decision.requires_approval:
return request_human_approval(decision)
return execute_with_policy(decision)
The exact API matters less than the ownership boundary. The gateway, not the prompt, owns feature flags, timeouts, retry policy, cost limits, approval, and audit events.
Prompts explain; code enforces
A prompt can tell the model not to perform an unsafe action. This is useful guidance, but it is not a control plane. Prompts can be misunderstood, overridden by context, or changed during iteration.
Deterministic enforcement makes the product contract testable. A denied capability remains denied even when the model asks persuasively.
Better failure semantics
Central governance also improves ordinary reliability. Tool timeouts, retries, circuit breaking, and idempotency do not belong in natural-language instructions. They belong in infrastructure that can observe outcomes and record them consistently.
The useful split is simple: models choose among possible intentions; runtimes decide which intentions may become effects.
Join the discussion
Comments are not open yet.