01Start with a testable outcome
Imagine an agent preparing a proposal from a customer request. It can search a catalogue, compare options, and draft a document. Before connecting those tools, define what completion means: a proposal that covers the requirements, uses prices from a current source, and identifies any unanswered questions.
That outcome should be checkable outside the model. Required fields and validation against the catalogue provide evidence. The statement “it is ready” does not, by itself, establish that the task is complete. The design should also distinguish between preparing a document and having permission to send it.
02Make limits part of the workflow
Give the agent a budget for time, tool calls, and cost. Enforce those limits in the code that coordinates execution. When a query repeatedly returns the same error, another attempt without new information can consume resources without bringing the task closer to completion.
A defined exit path preserves partial work and explains what remains. A missing price calls for another source or a review; an incomplete request calls for clarification. The right response depends on the obstacle. Designing these paths before deployment gives the surrounding system a way to handle each situation consistently.
03Evaluate the decision to stop
Useful evaluations include ordinary requests, conflicting data, unavailable tools, and insufficient permissions. Review both the final output and the actions taken. Record the tools used, relevant responses, and the reason execution ended, while avoiding unnecessary storage of sensitive information.
The aim is to distinguish three states: a verified result, work that requires intervention, and an execution halted by a limit. Communicating those states clearly lets someone continue the process without having to reconstruct the entire attempt.
| Situation | Next action |
|---|---|
| A price is missing from the catalogue | Flag the missing detail and request a review. |
| The proposal is complete | Present it to the person who approves sending. |
| The tool keeps failing | Stop retrying and save the partial work. |
References
Anthropic — Building effective agentsHave a process that needs a different way to work?
Let’s talk