Designing AI agents with clear boundaries

The word agent can make a workflow sound more capable than it is. In a product, an agent is software that can interpret a goal, choose from available tools and take one or more actions. That shift from suggesting to acting changes the design problem. Teams must decide what the system may do, what it must ask first and how a person can understand or reverse its work. Start with a narrow job Choose a task with a clear beginning, a small set of approved inputs and an observable finish. Drafting a support reply for review is easier to bound than resolving every customer issue. Preparing a deployment summary is easier to audit than allowing a system to change production infrastructure without a checkpoint. A narrow first task reveals the data, permissions and exceptions that a broad promise can hide. Give each action a policy List the tools the agent can use and the data each tool exposes. Separate reading from writing, low-impact actions from consequential ones, and reversible changes from actions that cannot easily be undone. Use the smallest permission that lets the workflow succeed. Require human approval when an action affects money, access, legal commitments, safety or production availability. Design for interruption and recovery Show what the system is doing, what information it used and what it plans to do next. Provide a pause or cancel path before a consequential step. Record tool calls and outcomes so an operator can investigate a surprising result. If a dependency fails or confidence is low, stop cleanly and explain what remains for a person to decide. Measure task quality, not activity Count completed tasks only when they meet an agreed quality bar. Track correction rate, escalation rate, time saved, permission failures and harmful actions caught before execution. Review real edge cases with the people who own the workflow. Expand autonomy only when evidence shows that quality, oversight and recovery remain dependable. An agent earns trust through bounded authority and predictable behaviour, not through a long list of tools.