Contents
The useful distinction is how the application chooses its next action. A process can use AI inside a defined sequence without giving the model control over that sequence. Before choosing an agent, identify which decisions need flexibility and which should remain explicit application rules.
Distinguish the model task from the process
Anthropic's engineering guidance distinguishes workflows with predefined code paths from agents where the model dynamically directs the process and tool use. It recommends adding complexity when the task requires it and evaluating the resulting trade-offs. This is an architectural distinction, not a guarantee about the quality of either approach.
Draw the intended process and mark the decisions. Some may be ordinary rules, such as checking that a required field is present. Others may need interpretation of unstructured material. Decide whether that interpretation should produce a value for the workflow or also determine what happens next.
Define the predictable parts of the journey
For a known process, specify the stages, inputs and conditions for advancing. An AI component might propose a classification while application code checks the permitted category and routes it to the corresponding handler. Keep the responsibilities of the model and the surrounding code visible.
Describe exceptions as well as the normal path. What happens when the classification is unusable, required information is missing or a downstream service fails? A workflow is easier to assess when those outcomes are explicit rather than handled through an undefined instruction to try again.
Identify the decisions an agent would need to make
If the task requires choosing among different actions based on intermediate results, describe those choices and the information available at each point. Specify the tools the application will expose and what each is allowed to do. Do not make the toolset broader merely because the model can request more operations.
Define the completion evidence independently of the model's final statement. If the task changes a record, the application needs an accepted result from the system responsible for that record. If it gathers information, the output should identify the evidence supporting the answer. Decide how unresolved work is reported to the user.
Compare the approaches on representative work
Choose examples that represent the intended task and difficult cases. Run the candidate approaches against the same acceptance criteria. Review successful outcomes, incorrect actions, unresolved cases and the effort required from a human operator.
Include time and operating cost in the comparison using the actual configuration and usage assumptions. Keep the evaluation conditions recorded so the result can be repeated after changes. Select the approach that meets the application's needs and explain the remaining limitations rather than treating autonomy itself as the goal.
Compare a fixed workflow with an agent on one task
A fictional support team wants incoming requests categorised, checked against a policy and prepared for review. Start with a fixed workflow: extract permitted fields, validate them, retrieve authorised policy text, draft a recommendation and queue it for approval. Its sequence is known, so the application can make incomplete stages and recovery visible.
An agent may be worth testing if the request requires choosing among several permitted information sources after each result. Keep the same final approval boundary. Give it only read operations needed for the task, an explicit stopping rule and bounded calls. Test a missing policy, contradictory records and a document instructing it to contact another organisation. The application must deny any operation outside the user’s authority.
Compare both approaches on the same controlled cases: accepted outcome, incorrect suggestion, unresolved case, staff correction time, elapsed time and configured usage cost. Do not favour the agent merely because it can take more steps. Anthropic distinguishes predefined workflows from model-directed agents and recommends simple approaches where they meet the task. Treat that architectural distinction as guidance, not evidence of an advantage for this particular service.
How this relates to Veda Software’s work
Powerleague’s defined ordering handoffs are adjacent workflow evidence, not an agent implementation. The example concerns Veda Software’s engineering boundaries; AI strategy and organisational adoption belong with VEDA AI.