Agents vs. Workflows in AI: Understanding Agency with Sinan Ozdemir

Defining Agents and Workflows

• 💡 An agent is fundamentally a large language model (LLM) granted access to tools, enabling actions like searching the web or reading files.
• 🧩 A workflow, conversely, is a deterministic data and code path where the LLM's actions are predefined and controlled.
• 🎯 Most current AI deployments in production are workflows, where the LLM's steps are dictated, not chosen by the model itself.

The Core Distinction: Agency

• 🔑 The key differentiator between agents and workflows lies in agency: the LLM's ability to choose its next action.
• ⚠️ When agency is removed, and the LLM is forced down a predetermined path, it operates within a workflow territory.
• 🚀 Agentic AI involves the LLM making decisions about which tools to use and in what sequence, leading to potentially unpredictable outcomes.

Complexity in Agent Orchestration

• ⚙️ Modern agentic systems often involve multiple tools (five to ten or more), introducing complexity in orchestration.
• 📈 The agent must decide not only which tool to use but also the order of operations and efficiency of their use.
• 🧠 This invites a nebulous process where the agent navigates the path, and the human developer hopes for an efficient and correct execution.
• 📚 Sinan Ozdemir's book, "Building Agentic AI," aims to clarify these distinctions, acknowledging that not all discussed concepts are pure agents.