Part 5 of the Things I Didn't Know About Claude series.
Two more things the quizzes caught me on. One about architecture. One about a default value.
Workflows Are Not Agents
The quiz asked about "a 4-step predefined process that enhances task execution." I picked "agent." Wrong. It's a workflow.
Anthropic's distinction is clean:
- Workflow: predefined steps, fixed control flow, predictable. Your code decides what happens next.
- Agent: autonomous tool use, dynamic control flow. The LLM decides what happens next.
I'd been calling everything an "agent." My prompt chains? Agents. My routing logic? Agent. My parallel processing pipelines? All agents.
They're workflows. Every one of them. An agent is specifically when the model itself decides the next action — like Claude Code in autonomous mode, where it reads code, decides to run tests, interprets results, and fixes bugs without a predetermined sequence.
The Five Workflow Patterns
Anthropic categorizes workflows into 5 patterns:
- Chaining — serial steps, output of one feeds the next
- Routing — classify input, route to specialized handler
- Parallelization — run multiple prompts simultaneously, combine results
- Orchestrator-workers — a coordinator delegates subtasks
- Evaluator-optimizer — generate, evaluate, improve in a loop
Each has fixed control flow designed by the developer. The course showed a concrete example: an image-to-3D pipeline where Claude describes a photo, CADquery generates a model, a renderer creates an image, and Claude compares it to the original. Every step is predetermined. It's a workflow, not an agent, even though an LLM is involved at multiple points.
Why the Distinction Matters
The course's advice was direct: always check if a workflow can solve your problem before reaching for an agent. Agents have lower task completion rates. They're harder to test. They're harder to evaluate.
Workflows break big tasks into small, testable steps. Each step's accuracy improves because the task is focused. You can evaluate each step independently. You can swap out components.
Agents are for when you genuinely can't predetermine the control flow — when the task requires dynamic tool selection based on intermediate results. That's a smaller set of problems than most people think.
Temperature Defaults to 1.0
The other thing: Claude's default temperature is 1.0. Maximum creativity. I'd assumed something moderate — 0.5 maybe.
For factual extraction, classification, or data processing, temperature 1.0 is actively harmful. You want temperature: 0 for deterministic, consistent outputs.
I'd been running classification prompts at temperature 1.0 for months. The results were mostly fine, but occasionally inconsistent. Now I know why.
One more detail: when Extended Thinking is enabled, temperature is locked at 1.0. You can't change it. The thinking process needs the full probability distribution.
This is the last topic-specific article in the series. Next: a summary of all the things Anthropic Academy's quizzes taught me.
Anthropic Academy is free: anthropic.skilljar.com
Related:
Is your Claude Code setup actually safe? Run npx cc-health-check — a free 20-point diagnostic. Score below 80? The Claude Code Ops Kit fixes everything in one command.
How do you draw the line between workflows and agents in your projects?
Top comments (0)