What This Is
A long technical piece on Juejin makes the point bluntly: for tasks like fab (wafer plant) anomaly handling and invoice recognition — tasks that look "system-complex" — the next action is already written into the process diagram. No Agent is needed. Most so-called Agent projects flooding the market are simply Workflows dressed up in an LLM coat.
The judging rule is a single question: who decides the next step. If the developer pre-draws the flow, it's a Workflow; only when the next step can only be decided after evidence comes in do you need an Agent.
The author adds a second, more usable test — the LLM removal test: pull out the large language model, and see whether the entire process can still be drawn as a standard flowchart. If it can, the thing is a Workflow at heart.
The cost is the real point. Agents trade flexibility for lower predictability, harder auditing, and higher cost. Forcing an Agent onto these tasks introduces unnecessary risk onto the production line.
What does a genuine Agent task look like? Insurance claim investigation, cloud cost anomaly hunting — the goal is clear, but the investigation path only becomes clear when the previous step's findings come in. You can't write a single line of the flow before the work starts.
Industry View
The technical community broadly welcomes this myth-stripping framework. Over the past year, the Agent concept has been seriously inflated — many SaaS vendors slap "Agent" on any automation that touches an LLM, making it hard for decision-makers to tell the real thing from the wrapper. Having an explicit judging criterion at least gives the conversation a foundation.
But dissent is real. Some architects argue the binary is too clean: real systems usually sit somewhere on the spectrum. In multi-Agent systems, pre-set orchestration coexists with runtime decisions; forcing teams into one bucket just pushes them to draw wrong flowcharts. Others point out that "can you write it out" is itself an engineering judgment — and the traditional industries that need this judgment most are precisely the ones that lack it. Worse, the consultants best placed to help you judge are often the ones most eager to sell you Agents.
The more pragmatic risk: by this standard, a large share of in-enterprise AI automation projects should be reverted to Workflow. That means Agent budgets already spent and Agent engineers already hired will need to be re-examined. That is why this piece will not be liked by every reader.
Impact on Regular People
For enterprise IT: the next time a vendor pitches an "Agent platform," make them draw the process first. If they can draw it cleanly, negotiate the Agent portion off the price — you are paying for flexibility you do not need.
For individual careers: to judge whether your job will be replaced by an Agent, look not at complexity but at whether the next step is only knowable after doing it. Bookkeeping, typesetting, report reconciliation — paths are clear, this is Workflow land. Accident investigation, complaint root-tracing, solution design — paths are open, these are Agent targets.
For the consumer market: most "AI Agent" assistant products on the market are Workflow + LLM combinations. Their marginal cost will keep falling, but real Agent products carry high inference costs and won't get cheap in the short term. Don't pay a premium for "Agent" by itself.