Before letting an AI improvise, Anthropic says try a plain prompt, then a fixed script — only reach for a flexible agent when neither works.
What are the named workflow patterns Anthropic identifies before reaching for agents?
This explores which simpler, more predictable LLM setups Anthropic tells builders to try before they hand control to a fully autonomous agent, and why it puts them in that order.
This explores the step-by-step ladder Anthropic recommends climbing before building a full agent. One thing to know up front: the notes retrieved here describe the principle behind the ladder, but none of them lists the named patterns. The names below come from Anthropic's original "Building Effective Agents" post, which the first note summarizes. They are not drawn from the note text itself. The principle is to start with the simplest LLM call that works. Next, use workflows, meaning fixed sequences of LLM calls that you write in code, for any task that splits cleanly into steps. Use an agent only when you can't know the path in advance. The reason is cost: agents give you flexibility in exchange for latency and money, and that trade only makes sense when a fixed path won't work When should teams choose workflows over agents?.
In that post, Anthropic names five workflow patterns. **Prompt chaining** passes the output of one call into the next, sometimes with a code check between steps. **Routing** classifies an input and sends it to a specialized prompt or model. **Parallelization** comes in two forms: split a task into independent pieces, or run the same task several times and vote on the answer. **Orchestrator-workers** has a central LLM break a task into subtasks on the fly and hand each one to a worker. **Evaluator-optimizer** puts one LLM's draft through a loop where another LLM critiques it and the first revises. Read in that order, each pattern gives the model a bit more control over what happens next. Orchestrator-workers is the point where the workflow is close to becoming an agent.
The part that's easy to miss is what changes once you do build an agent. A workflow is safe largely because its structure is fixed. An agent loses that, so Anthropic adds two guardrails to replace it. First, every step has to check real results from the environment (test output, tool returns) rather than trusting the model's own account of progress. Second, every agent needs a hard stopping rule, such as a maximum number of iterations. Anthropic presents these as necessary safeguards, but it doesn't report any measurements of how well they work Can environmental feedback and iteration limits prevent agent errors?. Fixed structure also has a downside. In planner-executor systems, where one LLM writes the workflow and others run it, a carefully written prompt can bias the plan itself before any downstream check runs. In one study this raised the attack success rate by up to 55% Can prompts alone reshape multi-agent workflows without system access?. A workflow that an LLM generates dynamically is therefore a different kind of thing from one you hard-code.
The consulting firms approach the same question at the scale of a whole company, and they reach a related conclusion. Deloitte argues that agentic projects fail when agents are added on top of existing workflows. In its view, the process has to be redesigned from start to finish Does agentic AI succeed through process redesign or workflow layering?. McKinsey describes "agent squads" with humans supervising from above the loop, but notes that governance is behind deployment Can AI agents truly operate above the loop at enterprise scale?. Anthropic's own data adds a twist. Claude Code's longest sessions nearly doubled, to more than 45 minutes, in three months, which suggests models can already handle more autonomy than people give them Are AI models held back from their autonomy potential?. The "start simple" advice is about engineering discipline, not a claim that models can't do more. Many teams may be under-using agents, while others build them too early.