When OpenAI announced AgentKit, they called it a tool for building agents. But looking at their examples, what they're showing are deterministic workflows with predetermined steps—not true agents. This matters because agents and workflows are fundamentally different patterns for solving different problems.

In this video, I break down:
- What actually makes something an agent vs. a workflow
- Why Anthropic's definition (from their "Building Effective Agents" article) has become the standard
- The key difference: who decides when to stop—the LLM or your code?
- When to use agents (unclear paths, need for improvisation)
- When to use workflows (known solutions, need for optimization)
- Why most real systems exist on a spectrum between the two

Agents are like jazz—improvisation and feel. Workflows are like classical music—optimized upfront for perfect execution. Both have their place, but mixing up the terminology makes it harder to communicate about the trade-offs.

**Timestamps:**
[Add your own timestamps here]

**Resources mentioned:**
- Anthropic's "Building Effective Agents" article: https://www.anthropic.com/engineering/building-effective-agents

Become an AI Hero:

https://aihero.dev

Follow Matt on Twitter

https://twitter.com/mattpocockuk

Matt Pocock

OpenAI's new Agent Kit sparked debate about the definition of AI agents versus workflows. An agent is a loop where the LLM decides when to stop, calling tools iteratively to gain new information. A workflow uses predetermined steps with known code paths. Agents excel at improvisation and unclear solution paths (like coding assistants), while workflows are better for repetitive tasks with optimization opportunities. Most real-world systems exist on a spectrum between these patterns, combining both approaches. The distinction matters for understanding trade-offs and communicating system design effectively.

OpenAI just confused everyone again

<p>Wonderful!  we’ve invented code that can now decide when to be confused instead of just failing silently.</p>
