Hacker News Reader

Comment by andy99 | Hacker News Reader

Anthropic has a definition:

  Workflows are systems where  LLMs and tools are orchestrated through predefined code paths.
  Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks

https://www.anthropic.com/engineering/building-effective-age...

While I know it's a marketing term, I think a good distinction is that agents have a loop in the execution graph, and can choose whether to loop or keep going. Workflows are chained LLM calls where the LLM has no "choice".

View on HN

I should have read this 12h ago! This afternoon, I tried to create my first simple agent using LangChain. My aim was to repeatedly run a specific python analysis function and perform a binary search to find the optimal result, then compile the results into a markdown report and export it as a PDF.

However, I now realize that most of these steps don't require AI at all, let alone agents. I wrote the full algorithm (including the binary search!) in natural language for the LLM. And although it sometimes worked, the model often misunderstood and produced random errors out of the blue.

I now realize that this is not what agents are for. This problem didn't require any agentic behavior. It was just a fixed workflow, with one single AI step (generating a markdown report text).

Oh well, nothing wrong with learning the hard way.

That reminds me of another recent submission that seems relevant:

"Don’t let an LLM make decisions or execute business logic"

319 points, 168 comments, 1 day ago - https://news.ycombinator.com/item?id=43542259

And the LangChain definition is a further simplification of Anthropic's:

"An AI agent is a system that uses an LLM to decide the control flow of an application."

What gets left unsaid is whether current SOTA LLMs actually have the reasoning and planning capabilities to do this reliably. I would argue that except for code debugging tasks and simple research tasks (iterative Googling with a web browser and then formatting the results into a report format) they do not. That may change in 6 months, but right now the hype has gotten ahead of the capability.

Reasoning, planning, and reliability do not seem to be strong features of current LLMs.

> Workflows are systems where LLMs and tools are orchestrated through predefined code paths

This definition keeps coming up, but the definition isn't accurate for workflows. Modern workflow systems are very dynamic in nature and direct their own process and tool usage (e.g. like Temporal, disclaimer: my employer). You can even write workflows that eval code if you want though for most that's a step of flexibility too far to give to an LLM. Many workflows have LLMs tell them what to do next, sometimes via a bounded tool list, or sometimes open ended e.g. process execution or code eval. There is no limit here. A better definition of a workflow is that it durably orchestrates things, not that the sequence or code is predefined.

So by a more accepted/modern definition of "workflow", agents are workflows that just happen to be more dynamic than some more rigid workflows.

But they muddle it by saying all of them are Agentic Systems

Agentic vs Workflow boils down to age old computing paradigm of Declarative vs Imperative.

Seems like more of a special case than a different thing altogether

the LLM does have the ability to send garbage output in protest

Reply on news.ycombinator.com