Claude GuideDevantia × Executive Partners Group
DevantiaExecutive Partners Group
0/23 Jean-Christophe Leroy
Jean-Christophe Leroy
Guide author
✓ Verified on September 3, 2026⏱ 5 min read
Module 15 bis

Workflow or agent? The 5 patterns that work

In plain terms — the question to ask before you build The previous module shows you how to build an agent. This one answers the question that comes just before it: do you actually need one? In most cases, no — and that comes from Anthropic itself. Here is the decision framework and the five proven architectures, so you don't pay for a Ferrari where a bicycle would do.

The golden rule: start simple

Anthropic publishes a recommendation few vendors are willing to make: don't start with an agent. Agentic systems "trade latency and cost for better task performance". For most uses, a single well-designed call — with the right examples and the right documents in context — is more than enough.

⚠️ The reflex that costs you Only add complexity when it objectively improves a result you are measuring. A setup that takes four minutes and costs $2 where a well-written prompt takes eight seconds and costs two cents isn't modernization: it's a regression dressed up as innovation.

Two families, not one

🛤️

The workflow

Claude and your tools follow a path you mapped out in advance. Predictable, testable, budgetable. Covers the bulk of enterprise needs.

🧭

The agent

Claude chooses its own route and its own tools, and keeps control until the job is done. Powerful on open-ended problems, unpredictable by nature.

Three questions before you build YOUR TASK 1 · Would one well-written prompt do the job? YES Simple prompt + examples + context NO 2 · Are the steps known in advance? YES WORKFLOW — the 5 patterns predictable · testable · budgetable NO AUTONOMOUS AGENT + guardrails, always 3 · Step back down a level whenever the measured result stops improving

The 5 workflow patterns

These are the five architectures Anthropic documents and publishes as open source. Each answers a specific situation — the skill lies in recognizing which one you are looking at.

1

Prompt chaining — the assembly line. Each step processes the output of the previous one, in an order you set.
When: the task breaks down into sub-tasks known in advance.
Example: draft the product pitch → translate it into three languages → adapt it to each channel's format. Every step is verifiable; if quality slips, you know exactly where.

2

Routing — the switchboard. A first call classifies the request, then sends it to the right treatment.
When: your inbound falls into distinct categories that call for different answers.
Example: customer requests sorted between general question, refund request and technical incident — each with its own prompt and tools. Budget bonus: simple cases go to Haiku, heavy ones to Opus 5.

3

Parallelization — the reading committee. Two variants: sectioning (each handles a portion) and voting (several handle the same thing, you compare).
When: you are after speed, or confidence through cross-checking.
Example: a contract reviewed simultaneously from the legal, financial and operational angle — three parallel readings, one synthesis.

4

Orchestrator and workers — the project manager. A central Claude breaks the mission down on the fly, delegates, then consolidates.
When: the mission is complex and how it splits depends on what you find along the way. That is the difference from chaining: here the sub-tasks are not predictable.
Example: the acquisition due diligence described in module 15.

5

Evaluator-optimizer — the writer-editor duo. One Claude produces, another critiques, and the loop runs until the target level is reached.
When: you can state clear quality criteria and iteration visibly improves the result.
Example: a commercial proposal reworked until it passes your checklist — hook, hard evidence, objections handled, call to action.

So when does the autonomous agent apply?

It starts from your instruction, then plans and operates on its own. It relies on feedback from its environment — tool results, errors it hits — to judge its progress, and can come back to you when it lacks information.

Save it for open-ended problems whose steps you cannot predict, in a trusted environment and always with human checkpoints. Resolving a complex technical ticket, driving a workstation: yes. Processing your expense claims: no, that's a workflow.

📌 Anthropic's three principles Simplicity: the leanest design that does the job. Transparency: show the planning steps explicitly — an agent that works as a black box is an agent nobody will dare put into production. Rigor on tools: document and test every tool you hand it as carefully as you do your prompts.
🎯 What to take to your board When a vendor offers you "an AI agent", ask three questions: are the steps predictable? (then it's a workflow — cheaper and more reliable), what is being measured? (without a metric, the complexity cannot be justified), where are the human checkpoints? (without them, you will never deploy).
📚 For your technical teams Anthropic publishes the reference implementation of these five patterns as open source, in ready-to-run Python notebooks: claude-cookbooks / patterns / agents. The founding article, Building Effective Agents (Erik Schluntz & Barry Zhang), takes twenty minutes to read and is worth it even without a technical background.
🛠️ Your turn — 5 minutes

Facing an automation need, answer the module's three questions: is one prompt enough? are the steps known? what is measured?

Expected result : You can say whether it is a prompt, a workflow or an agent, and why.