Engineer for reliability with agent boundaries

The goal isn't an agent that can do anything. It is an agent that does the right thing inside a well-defined box, and reliably refuses or escalates outside it. Reliability is largely a function of constraints.

Constraints that make behaviour stable

• Explicit planning and step budgets: cap tool calls, runtime and retries.
• Typed interfaces: tools take structured inputs; outputs are validated.
• A policy engine: rules checked before execution.
• Approval gates for irreversible or high-impact actions.
• Fallback modes: degrade gracefully to recommend instead of act.

A reliable pattern

The agent does the reasoning and orchestration; a workflow engine provides the execution rails. Don't let a model freehand critical actions when deterministic workflows can enforce policy and correctness.

The secret to safe agents is not better prompts. It's system design.