Decision Models Make Bounded Judgment Cheaper

Decision models make bounded judgment cheaper. The boundary is still ours to define. TypeSafe introduced Jev: structured, probabilistic outputs rather than generated text. Since then, Databricks, AWS, OpenAI and Microsoft have introduced related decision-oriented capabilities.

An LLM can propose a tool call. Whether that call fits the workflow is a separate question. Authorization must remain a policy decision. A decision model gives that judgment a typed interface: Is the call grounded in the available context? Should the workflow retry or escalate? It returns a bounded answer, often with a score, rather than generated prose. Hooks can already intercept calls before execution. The model doesn't invent the gate. It makes a specialized check more practical to run there.

The enterprise standard should be the control contract, not a decision model everywhere. The runtime can assess workflow fit. A tool gateway can enforce access. Some actions need both. The agent cannot skip required checks. This is a useful new primitive for bounded autonomy.

The engineering remains ours: record the proposed action, the model's result, the policy version, and the resulting disposition. Name the policy owner. A score alone gives you neither accountability nor a defensible audit trail.

Low confidence can trigger clarification or human review. A prohibition doesn't become negotiable because the model is confident. Validate scores against outcomes in your workflow. A typed answer can be wrong. The control can live inside the runtime. It must stay outside the model's discretion.

For the wider control-plane argument — identity, gateway, registry, observability — see Enterprise Agent Architecture.