Anthropic told builders to try workflows before autonomous agents

The conversation

On 19 December 2024, Anthropic published an engineering essay, "Building Effective Agents," arguing that most systems calling themselves agentic should be built as one of a small set of named patterns rather than as an open-ended autonomous agent. The essay states that "some customers define agents as fully autonomous systems," while "others use the term to describe more prescriptive implementations that follow predefined workflows," and that Anthropic groups both under "agentic systems" while drawing its own line between the two.

What the documents show

Anthropic's essay defines workflows as "systems where LLMs and tools are orchestrated through predefined code paths," and reserves "agents" for systems where the model "dynamically direct their own processes and tool usage." It names five workflow patterns directly, prompt chaining, routing, parallelization, orchestrator-worker and evaluator-optimizer, each with a stated use case, and points to a companion tool-use guide for the mechanics either a workflow or an agent needs to call functions at all. The essay also names third-party frameworks its authors have seen builders use for these patterns, including Rivet, described as "a drag and drop GUI LLM workflow builder," and Vellum, without endorsing either.

The system boundary

The essay's recommendation is a boundary on complexity, not capability: it states that "workflows offer predictability and consistency for well-defined tasks, whereas agents are the better option when flexibility and model-driven decision-making are needed at scale," language that puts the choice on the builder rather than treating agent autonomy as a strictly better default. Anthropic frames this as its own engineering guidance, drawn from patterns it says it has "seen emerge across various customer deployments," not as a specification or standard that other vendors' documentation is bound by.

Where it fails

Because the essay is Anthropic's own vendor guidance rather than a benchmark or an audited study, it does not claim that any of its five patterns reduces error rates or cost by a stated amount, and a reader should not import such a figure from it. Its distinction between "workflow" and "agent" is also a definition Anthropic chose, not one every framework or vendor in this record uses identically, so a claim that a given product "is an agent" should be checked against whatever that vendor's own documentation means by the word.

• Does a system described as agentic match Anthropic's workflow definition, its agent definition, or neither?
• Which of the five named patterns, if any, does a given deployment's own documentation claim to use?
• Is a claimed benefit of using agents sourced to the vendor's own guidance, or to an independent measurement?

The essay's most durable claim is negative: it recommends starting with the simplest workflow that solves the problem and reaching for full agent autonomy only when a workflow cannot, a sequencing this desk treats as one vendor's engineering judgment worth naming as such.