Agentic Workflow

An agentic workflow is a software process in which one or more Large Language Models (LLMs) use tools, intermediate state, and multiple model calls to complete a task. It combines probabilistic model decisions with conventional application logic.

The term covers a spectrum of systems. At one end are deterministic workflows where code defines every transition. At the other are autonomous AI agents that dynamically choose their next action. Anthropic's commonly used distinction is that workflows follow predefined code paths, while agents direct their own processes and tool usage.

Common agentic workflow patterns include:

• prompt chaining, where a task is decomposed into sequential model calls;
• routing, where a classifier or model sends requests to specialized prompts, models, or tools;
• parallelization, where independent subtasks run concurrently and their outputs are combined;
• evaluator-optimizer loops, where one model generates work and another critiques it;
• orchestrator-worker systems, where a central model delegates subtasks; and
• autonomous action loops, where an agent observes, plans, acts, and evaluates progress.

An agentic workflow is useful when a task cannot be handled reliably by one prompt, one retrieval step, or one model response. It can improve quality through specialization, verification, and iteration, but typically increases latency, cost, and operational complexity.

The architecture should make state transitions explicit. Each step should have a defined input, expected output, retry policy, timeout, and completion condition. Structured output is often used at workflow boundaries, while tool calling connects model decisions to external systems.

Workflow designers should also account for nondeterminism. A local failure can propagate through subsequent steps, and unrestricted loops can consume substantial resources without making progress. Checkpoints, maximum-step limits, human approvals, and agent evaluation are therefore important production controls.

Agent orchestration is the narrower discipline concerned with coordinating the components of these workflows. Anthropic provides an overview of production patterns in Building Effective AI Agents.