Workflow or agent?
A workflow follows predefined steps. An agent can choose its steps and tools dynamically. This distinction, explained by Anthropic, helps guide architecture choices. A stable invoice approval process does not necessarily need substantial autonomy.
Define what it may do
Our recommendation is to begin with a limited task, such as preparing a case file from authorized documents. Separate reading capabilities from actions that change data or commit the organization. Set a budget, a maximum number of steps and situations requiring human approval.
Documents read by an agent may contain malicious instructions. Anthropic’s work on trustworthy agents describes this prompt-injection risk. Document content should not become authorization to act.
Measure the task actually completed
A convincing answer does not prove an action succeeded. For a request-processing agent, check the created record, completed fields and unintended changes. Compare several trials because behaviour can vary.
We advise measuring total effort, including human review and correction. Gains depend on the use case, data quality and cost of errors. No universal productivity percentage can be promised before that evaluation.
Sources and features may change. Check the current version when planning your project.
All articles ↗