New term: agents vs workflows
2024-12-19 · Publisher : Anthropic · Domain : Artificial Intelligence
Anthropic separates workflows, where code chains model calls along predefined paths, from agents, where the model decides its own steps and tools. The rule: start with the simplest solution.
What changes
- A shared vocabulary for who holds the thread: the code or the model.
- Five workflow patterns become references: prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer.
- Teams are encouraged to skip heavy frameworks until they are needed.
- An autonomous agent becomes a choice to justify, not a starting point.
### Definition In Anthropic's vocabulary, **workflows** are "systems where LLMs and tools are orchestrated through predefined code paths", while **agents** are "systems where LLMs dynamically direct their own processes and tool usage". Both fall under "agentic systems". ### Origin Erik Schluntz and Barry Zhang published "Building effective agents" on Anthropic's website on December 19, 2024, drawing on work with dozens of teams. Their core advice: "Start with simple prompts, optimize them with comprehensive evaluation, and add multi-step agentic systems only when simpler solutions fall short." ### What the term covers - **Prompt chaining**: a task split into successive steps. - **Routing**: an input is classified, then sent to the right handler. - **Parallelization**, **orchestrator-workers** and **evaluator-optimizer** (one model produces, another critiques). - Agents fit open-ended problems whose steps cannot be predicted, such as some software development tasks. ### Criticism and limits The line is practical rather than theoretical: many real systems mix fixed steps with moments of autonomy. In marketing, "agent" is still used far more loosely. ### Key takeaway Before building an agent, ask whether a workflow is enough: it is more predictable, cheaper and easier to test.
Links
Tags : Vocabulaire, Agents, Workflows, Architecture