## Atomic Insight OpenAI draws the workflow/agent line at control, not capability. A workflow is "a sequence of steps that must be executed to meet the user's goal" — an application can integrate LLMs and still not be an agent if it doesn't let the model control execution (a simple chatbot is OpenAI's example). An agent, by contrast, "leverages an LLM to manage workflow execution, make decisions, and correct its actions if needed," dynamically selecting the right tools to gather context or act, based on the workflow's state, always within guardrails. That control test doubles as OpenAI's build criteria: reach for an agent when a task involves complex decision-making (nuanced judgment, exceptions, context-sensitive calls like refund approval), difficult-to-maintain rules (rulesets so extensive that updates become costly or error-prone), or heavy reliance on unstructured data (interpreting natural language, extracting meaning from documents, conversational interaction — e.g. processing a home insurance claim). Absent those, OpenAI's implication is a deterministic workflow suffices. **Cross-Domain Connections**: [[Anthropic's Agentic Systems Umbrella Spans Workflows to Agents]], [[OpenAI's Three Core Agent Components]] ## Source - [[How to Build AI Agents]] — infographic by Jaynit Makwana (@JaynitMakwana) on X, summarizing OpenAI's published agent-building guidance, saved 2025-06-18