olivlaw

From books

First a workflow with fixed steps; an autonomous agent only when the steps cannot be known in advance.

Erik Schluntz, Barry Zhang · Building effective agents · 2024 · Building effective agents (Anthropic, 19 decembrie 2024) — introducere1 minute read
“Consistently, the most successful implementations use simple, composable patterns rather than complex frameworks.”Erik Schluntz, Barry Zhang · Building effective agents · 2024 · Building effective agents (Anthropic, 19 decembrie 2024) — introducere

If you can draw the flowchart, build a workflow, not an agent.

The text separates workflows (model and tools orchestrated through predefined code paths) from agents (the model directs its own process). Five workflow patterns: prompt chaining, routing, parallelisation, orchestrator–workers, evaluator–optimiser. The decision rule: if you can draw the flowchart, build the workflow — it is cheaper, testable, predictable. Move to an agent when the number of steps is unknown and the task needs the model's judgement at every step (open-ended coding, research). The agent's price: latency, spend, compounding errors. Start with the simplest thing that works and add autonomy only when measured evaluations say the workflow falls short.

The flowchart is more than an eligibility test: it is a design commitment. When you can draw it, every decision the system makes becomes a box on paper, and therefore a box you can put a checkmark on. That changes how you debug. When a result is wrong, you don't ask "what was the model thinking?" but "which box did it go wrong in?". Think of a return form in an online shop: if the item is defective you go down path A, if you simply didn't like it, path B. When something fails, you know exactly which branch to fix. With an autonomous agent, the mistake can appear anywhere in a chain you never drew. So the next question isn't "how smart is the model?", but "can I afford to lose visibility into its decisions?".

Passage 2: There is another, subtler boundary between workflow and agent: the cost of errors. In a workflow, an error stops where it is caught and can be fixed locally. In an agent, steps lean on each other, so a small early mistake can grow with every step. Think of cooking from a fixed recipe: if you add too little salt, you correct it at the next step. But if you improvise without a recipe and slip early, every later decision builds on the wrong foundation. That is why autonomy costs not just money, but reliability. The practical question becomes: how quickly would my system notice that the first step was wrong?

Why it matters Most "agents" in production are well-named workflows; they are also the ones that do not fall over.

A singlemodel callPromptchainingRouting,parallelisationOrchestrator–workersAutonomousagent
Climb one rung only when the evaluations say the rung below falls short.
Open on YouTube ↗

Shelves

← Back to the feed