Automation and agents address different problems
In a predefined automated workflow, the steps and conditions are specified in advance. With an agent, the model can help choose actions during execution. Anthropic uses this distinction to separate orchestrated workflows from systems with greater autonomy.
There is no benefit in adding AI decision-making to a rule that is already clear. Receiving a valid form and creating a record may require a simple integration. Interpreting an ambiguous request and retrieving information from different sources may justify a different design.
Define the task before granting access
Fictional example: an assistant consults the help center and drafts a response to a support ticket. It can search and summarize information; granting a discount or changing a contract are different actions that need not share the same authorization.
- Objective
- Describe the expected output and what is outside the scope. A defined task makes the result easier to evaluate.
- Tools
- Provide only the information and actions the task needs. Distinguish lookups, draft preparation and changes to records.
- Stopping point
- Plan for uncertainty, conflicting sources, missing information and out-of-scope requests. Define who receives each case.
The demonstration needs to include what goes wrong
OpenAI provides tools and monitoring capabilities for building agents. These are part of implementation; they do not replace process design or establish that an agent works well in a particular situation.
Test incomplete inputs, unexpected tool results and attempts to move beyond the task. Also check recovery: can the team understand what happened and continue without repeating an action that has already been performed?
Start where the result can be checked
We recommend starting with a task whose result can be verified before it has external effects. Compare execution effort, review time and observed failures. Expand autonomy only when there is evidence to support that decision.
At A1, the process assessment defines the opportunity and its boundaries. Some processes will call for an agent; others will need a clearer rule, an integration or better-distributed responsibility.
