An AI assistant that suggests an action and an agent that performs it create different management problems. A suggestion can be ignored. An executed refund, database update or email has consequences outside the conversation. As teams experiment with more autonomous systems, the important design question is how authority travels from a person to a piece of software.
Separate a plan from permission
Anthropic’s published guidance on effective agents discusses composable workflows and agent patterns. It is useful background, rather than proof that any particular architecture fits every organization. A practical starting point is to separate reading information, proposing a change and applying a change into distinct capabilities with different permissions.
In a hypothetical purchasing workflow, an agent might compare suppliers and prepare an order without being allowed to submit it. A manager could approve a defined spending limit and supplier list. The resulting permission should apply to that concrete action, rather than becoming an indefinite authorization for unrelated purchases later.
Make the stopping conditions visible
A bounded system needs more than a button labeled cancel. It should have limits on spending, elapsed time, retries and the number of external actions. It should also preserve an activity record that lets an operator distinguish an unsuccessful attempt from a completed action whose confirmation was lost.
That distinction matters during recovery. Repeating a request after a timeout can create duplicate payments or duplicate messages. Systems should use transaction identifiers and check observed state before retrying actions with external effects.
Design for a graceful handoff
Human review works best when the reviewer receives the evidence, proposed action and unresolved uncertainty together. Asking someone to approve an unexplained recommendation merely moves the ambiguity to another person. The next useful agent may be the one that knows precisely when its authority ends.
