Appropriate autonomy
An AI agent can interpret a goal, plan intermediate steps, use tools, and act across a sequence rather than produce a single response. That makes agents a powerful component of operational software. It also makes the boundary of the task more important.
The useful question is not how autonomous the agent can become. It is how much autonomy the workflow can responsibly support.
Begin with a bounded responsibility
Agents belong where the task has a clear objective, an authorized set of information and tools, observable actions, and defined completion or handoff conditions. Good candidates often contain repeated coordination work: collecting approved information, preparing a structured draft, checking conditions, routing an item, creating a record, or monitoring for an exception.
IBM distinguishes agentic systems by their ability to reason through problems, plan actions, and run multi-step tasks, while also emphasizing human oversight in organizational use. Oversight should be designed into the task rather than added as a generic instruction.
Match autonomy to consequence
An agent that gathers public information creates a different consequence from an agent that modifies customer data, communicates a legal position, changes infrastructure, or approves a financial transaction. The task boundary should account for reversibility, sensitivity, evidence quality, delay, and the harm created by a wrong result.
Useful patterns include drafting for approval, guarded actions after checks pass, normal-path automation with exception escalation, supervised multi-step sequences, and reversible low-consequence actions.
Give the agent fewer tools—and clearer tools
An agent should receive only the tools required for its bounded task, and each tool should enforce its own permissions and validation. A well-designed system does not depend on a model remembering every security policy.
When an external dependency fails, the agent needs an explicit outcome: retry within a limit, request missing information, select a safe fallback, or escalate. Improvisation is not a failure strategy.
Make uncertainty visible
Interfaces should reveal the request the agent understood, the sources it used, the actions it took, what it could not resolve, and the point at which a person is expected to decide.
The OECD AI Principles emphasize transparency and explainability, robustness, security and safety, and accountability. In practice, the relevant user needs enough visibility to understand the agent’s role, inspect important evidence, and intervene.
Know where an agent does not belong
An agent is a poor fit when the objective is ambiguous, the required information is unavailable, the decision relies on unrecorded judgment, the actions are difficult to reverse, or no one can meaningfully review the result. It is also a poor fit when ordinary software can perform the task more reliably with fewer dependencies.
The strongest agent systems are defined by appropriate autonomy: a bounded role, controlled tools, visible evidence, tested failure behavior, meaningful oversight, and an accountable owner.