“Let’s build an AI agent” is the default ask in 2026. But most of what gets sold as an agent should be something simpler, cheaper, and more predictable — a fixed workflow, not an autonomous one. Knowing the difference is the call that decides your cost, your risk, and whether the thing actually ships. This is the plain-language version of Anthropic’s engineering guidance.
Both use AI. The difference is who decides the steps — and it changes everything about cost and control.
The steps are fixed in advance. AI does the thinking inside each step, but the path from input to output is one you designed and can predict.
Like a kitchen following a set recipe — same dish, every time.
You give a goal; the AI plans its own path, picks its own tools, checks its work, and keeps going until it judges the job done. Powerful — and less predictable.
Like a chef told “make me dinner” — flexible, but you can’t predict the bill.
Why it matters to you: a workflow has a bounded, forecastable cost and behaves the same way every run, so it’s easy to trust and audit. An agent trades that predictability for flexibility — more AI calls, a variable bill, and the risk that one wrong early step throws off everything after it. Neither is “better.” They’re tools for different jobs, and the expensive mistake is using an agent where a workflow would do.
The rule from Anthropic’s engineers is simple: use the least complex option that solves the problem. Each rung up buys flexibility and costs predictability. Start at the bottom.
One well-crafted instruction, maybe with examples. Astonishingly often, this is the whole answer. Cheapest, fastest, most predictable.
Chain a few prompts, route by category, run checks between steps. Predictable cost, auditable path. Most business automations live here.
A “manager” step that splits work across several AI calls and combines the results. More moving parts, still designed by you.
Hand over the planning. Reserve for open-ended jobs where you genuinely can’t script the steps in advance — and wrap it in guardrails.
Can you write down the steps in advance? Then it’s a workflow — build it that way and enjoy the predictable cost. Only when the steps genuinely depend on what the AI discovers along the way do you need an agent.
If you can draw the flow on a whiteboard, it’s a workflow. Agents are for when the path can’t be drawn in advance.
Agents make many more AI calls. If a workflow gets you 90% there at 10% of the cost, the workflow wins.
Agents are safe only where each step produces verifiable output — a test passes, a number reconciles — not where errors hide.
Autonomy needs guardrails: limits, human checkpoints, and the ability to undo. No harness, no agent.
If you’re unsure, you want a workflow. The most common and expensive failure in 2026 isn’t “our agent wasn’t smart enough” — it’s building an unpredictable agent for a job a fixed workflow would have done cheaply, reliably, and with an audit trail. Sophistication is not the goal. The job done reliably is the goal.
The full taxonomy for your engineers — the augmented LLM, the five workflow patterns, agents proper, and the three principles. The deep version of this briefing.
Open the leaf →Who runs these systems — one operator commanding AI agents, with human specialists in the loop. The role behind the workflow.
Read briefing →Don’t buy an agent because the word is in the air. Climb the ladder from the bottom, stop at the first rung that does the job, and reserve real autonomy — and its cost — for the problems that genuinely need it.