A workflow follows a predefined sequence. An agent can choose its next step using the current task and feedback. Use a workflow when the path is known; consider an agent when the path genuinely varies and its actions can be bounded and checked.
Two architectures, different trade-offs
Anthropic distinguishes workflows that orchestrate models and tools through predefined paths from agents that dynamically direct their process. That is a useful vocabulary for design reviews. It separates a product requirement from a marketing label and makes the source of flexibility explicit.
An invoice intake system might always classify a document, extract fields and ask a reviewer to approve the result. A research assistant may need to choose which source to inspect next depending on what it has found. Both can use language models. The second has a stronger reason to make the execution path dynamic.
Start with the simplest useful loop
Write down the goal, the available tools, the stopping condition and the evidence that indicates success. A task without a stopping condition can become an expensive conversation with no useful completion state. A tool without a narrow contract can turn an ordinary mistake into a broad operational problem.
For a research task, success might mean a concise answer with sources that support each material claim. For a code task, success might require a specific change and relevant checks. These definitions should be visible to the person using the system. They also provide a basis for evaluation rather than asking a model to judge its own confidence.
Tools are interfaces, not magic abilities
A tool description should explain inputs, outputs, permissions and failure cases. Prefer a focused operation over a general-purpose escape hatch when the task is predictable. Return enough information for the agent to understand the result without flooding the context with unrelated data.
Consider retries carefully. Reading a document again is different from submitting an order twice. Operations with side effects need a way to recognize repeated requests and report an existing result. Keep authorization outside the model's discretion: a confident sentence is not a permission grant.
Autonomy should be earned through evidence
Begin with draft or read-only behavior. Evaluate on representative tasks and inspect the failures. Expand the allowed actions only when the evaluation and recovery process support the change. A person should be able to interrupt execution and understand what has already happened.
Multi-agent designs can help when subtasks are separable, but they also introduce coordination and verification work. Before adding another agent, ask whether a clearer tool, a shorter context or a deterministic step solves the problem. More activity is not automatically more progress.
Questions, answered.
Does every AI product need agents?
No. A single model call, retrieval step or predefined workflow may be sufficient for a well-defined task.
When is agentic AI useful?
When the route to the goal depends on intermediate findings, and the system can constrain, observe and verify the actions taken.
Sources & further reading
Source-linked explanation and engineering perspective. Programme details and product documentation can change; consult the original sources for their current terms and capabilities.
Anthropic — Building effective agents ↗Anthropic — Writing effective tools for agents ↗