The Anthropic post reads like an architecture guide but is really a p…

The post says to start simple. Prefer workflows over agents. Add the loop only when the loop clearly earns itself.

I have built the same agent three times.

Each time I told myself I was starting with a workflow. Each time I ended up with a loop. Not because the loop was needed. Because the loop feels alive, and a workflow feels like a Python script with a Claude call in the middle. One of those two things is fundable and the other one you have to defend.

That was the thing I could not get past on the second read. The post is dressed as an architecture guide, but the actual thesis is temperamental. You reach for the agent because the agent feels alive. That is a feeling. Anthropic is politely telling you your feeling is expensive.

Cost me about six weeks on one of those three builds. Not a disaster - I got a system that would have taken me a fortnight of if-statements, and it took me longer than a fortnight to build the loop and longer again to trust it. You can do the arithmetic.

The taxonomy in the post is fine. Prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer, then the loop. Read them in order and each one is a version of the same question: how many of the branches do you write in code, and how many do you leave to the model? The workflow patterns write more branches. The agent pattern writes fewer. The advice is: write more, until you have measured that writing fewer helps.

The word measured is doing all the work in that sentence.

I have measured maybe two of the ten agent-adjacent things I have shipped in a way I would defend to another builder. Two out of ten is generous. Most of the time I am measuring by whether the demo lands and whether the customer says yes, which is measuring, but it is not the measuring the post is asking for. It is not “does the model’s extra decision-making beat a switch statement.” It is “does the room like it.”

That gap is the whole game.

There is a line near the end that most people scroll past: “we actually spent more time optimizing our tools than the overall prompt.” That is the punchline. Read the post again through that line. Every workflow pattern is a way to constrain what the tools do. Every agent pattern is a bet that the tools are good enough for the model to steer. The tool interface is not an implementation detail. It is the entire fight.

I have watched a smart builder spend three days on a system prompt and twenty minutes on tool descriptions, then wonder why the agent got lost. The parameter names were terrible. One docstring said “the target.” Target of what? Anyway. The prompt was gorgeous.

Here is the bit I actually want to sit with. The post recommends starting with the LLM API directly, not a framework, because frameworks obscure the underlying prompt and response. Fair. I would go further. The framework is what makes it possible to skip the tool-writing step, because the framework already wrote the tools, and the tools it wrote were fine for a demo and are the reason your agent falls over in month two. If your framework let you ship without ever writing a docstring for the model, you did not build an agent. You wrapped one.

None of this is news to anyone who has shipped a loop.

The bit that is news to me, this read, is how much of “which pattern do I use” is really a proxy for a temperamental question. Do I want to be in control at every step, or do I want to feel like I built something that is thinking. I lie to myself about which one I want. I suspect a lot of us do.

The workflow patterns give you control. The agent pattern gives you the feeling. The post is asking you to take control unless the feeling has actually paid for itself in dollars, or resolved tickets, or an eval that goes up when you swap the two approaches side by side.

Which brings me to the question I actually have.

If you have shipped both - a workflow-shaped version of a thing and an agent-shaped version of the same thing - and you measured them against each other honestly, did the agent win? Or did the agent draw, and you kept the agent because you liked it more?