We noticed that an Agent primer aimed at Java developers has quietly climbed Juejin's trending chart. It delivers a verdict that Agent vendors won't enjoy: 80% of products labeled "AI Agent" are essentially workflows — predefined step-by-step programs — wrapped in a thin layer of AI.
What this is
The article offers a restrained definition: an LLM Agent is a system in which a large model participates in decision-making, uses tools to interact with the environment, and advances toward a goal based on feedback. The model is just one piece — the program handles execution, tools connect to external information, and feedback determines whether to continue. If any of these four elements is missing, it's a workflow, not an Agent.
There is one iron rule for telling the real from the fake: is the next action dynamically chosen by the model based on task state and feedback? If the flow is drawn by a developer in a diagram, that's a workflow. If the system can reroute based on intermediate results, that's an Agent.
Industry view
What makes this worth our attention is that it punctures the current "Agent" label bubble. The context for the article's rise is that the term has been overused by Chinese AI vendors since 2024 — virtually every B2B AI product now claims to be "doing Agents." But the piece itself cuts through: "Workflows can also call models, branch, and retry; having tools and loops alone isn't enough to qualify as an Agent." Anthropic's "Building Effective Agents" essay from last year takes the same position, arguing most enterprise scenarios should use Workflows rather than Agents, because they're cheaper, easier to debug, and produce stable results.
Dissent exists. Some Agent vendors call this overly academic. Real industrial systems are typically hybrid — fixed flows handle approvals and compliance, Agents handle information gathering and preliminary judgment. "The model decides the next step" is a matter of degree, not either/or.
Impact on regular people
For enterprise IT: when a vendor pitches an "AI Agent," first ask "who decides the next step." If the model dynamically chooses based on feedback, it's an Agent — expensive, results unstable. If it's hardcoded into a flow diagram, it's a Workflow — cheap, easy to debug. Don't pay Agent prices for Workflow goods.
For individual careers: part of knowledge work is shifting into "how to outsource tasks to an AI that can make its own decisions." Not replacement — a new collaboration model. Whoever learns first to design task boundaries and acceptance criteria gets to use this leverage first.
For the consumer market: over the next year you'll see an avalanche of "AI assistant" and "intelligent agent" ads. The same standard applies: can it complete a task from a single spoken instruction, or can it only pick from a few preset options? The ones that can actually do things are expensive and unstable; the ones that just let you pick — don't be impressed by the new label.