As AI agents transition from read-only interfaces to active executors, the interface between what the model wants to do and what the system allows becomes the critical failure point. Natural language is inherently ambiguous. Prompt engineering alone cannot prevent hallucinated API calls or destructive side effects.
The Data Contract Boundary
In the HELM architecture, models do not execute code directly. Instead, they synthesize a structured specification representing their intended action. This specification acts as a formal contract between the stochastic model and the policy-based runtime.
The Proposal Phase
When an agent determines an action is necessary, it generates a proposal conforming to a strict schema. This schema defines the intended effect, the required parameters with type constraints, and the justification linking the action back to the original user request.
The Validation Phase
Before any execution occurs, the runtime validates the generated spec against the predefined Agent Contract:
- Schema conformance: does the proposal match the expected structure?
- Type safety: are the arguments the correct data types?
- Constraint checking: do the values fall within acceptable ranges?
Shifting from Prompt to Protocol
By treating generated specs as rigid contracts, the burden of correctness shifts from the prompt to the protocol. If a model hallucinates a parameter or attempts an unauthorized action, the spec validation fails immediately. The GeneratedSpec remains a proposal. It needs approval and a boundary check before any side effect can happen.