← Back to Concept Contrast

Concept Contrast Fable

Agents and large model applications

Conclusion first:One knows what it should do next, while the other continues the conversation through a fixed call chain. Treating them as the same thing means what you deliver is either over-engineered or fails at the critical moment.

One knows what it should do next, while the other continues the conversation through a fixed call chain. Treating them as the same thing means what you deliver is either over-engineered or fails at the critical moment.

Fable

Machines that walk and machines that talk

Woodware Town had two kinds of mechanical boxes, both taught "to listen to people" by the same skilled craftsman. The first sat at the counter: when a customer asked about wood prices or care methods, it consulted its ledger and organized a smooth reply. A good answer satisfied the customer; a wrong one was corrected by the shopkeeper with a follow-up remark.

A second kind of box was installed in the workshop. When a client asked for a desk, it not only replied, but also checked inventory, measured boards, dispatched sawyers, and asked the bookkeeper for a quote; when materials were short, it revised the plan. If it got one step wrong, the people after it all went wrong too. The workshop box also remembered unfinished items. If beech was missing today, it would check again tomorrow morning; if the bookkeeper had not returned a price, it would follow up. Once the counter box gave its answer, responsibility usually stopped there.

At first the new shopkeeper called both kinds of boxes "talking boxes," and when purchasing he only compared whose answers were prettier. As a result, after the counter box was placed in the workshop, it could only explain why elm should be sawed, but could not actually arrange the sawing; and the workshop box placed at the counter seemed bulky and expensive. The new shopkeeper's first fix was to add a long string of "if the customer wants to make a table, remind them to find a craftsman" to the counter box. It could remind, but still could not truly arrange, follow up, and close out.

The old master had him look at the chain of responsibility: some boxes are responsible for making things clear, and some boxes are responsible for getting things done. The former mainly revolves around Q&A and content; the latter needs goals, steps, tools, permissions, and receipts. The old master also had him look at the failure ledger. If a counter box makes a mistake, it is mostly a matter of revising one explanation; if a workshop box makes a mistake, it may already have consumed lumber, delayed the schedule, and sent out a quote.

The shopkeeper then tried removing permissions from the workshop box, letting it only suggest plans. It immediately became as light as the counter box, but it could no longer deliver finished products. Once capability, permission, and responsibility are separated, the category changes.

Later, the town bought mechanism boxes no longer just by looking at sample answers. The purchase order had to specify: whether it could invoke processes, whether it could save state, whether it could roll back on failure, and whether a human reviewed it. Talking and doing are budgeted differently.

From then on the town no longer asked only 'is it the same set of craftsman techniques'. It also asked: does it merely answer, or will it act for the result; when it errs, does the impact stop at one sentence, or does it bring down the whole production line.

Agent

An execution unit with environment awareness, autonomous planning, tool calling, and result verification capabilities. It can have multiple steps, multiple rounds of decision-making, and even change strategy midway through a task.

Large model applications

An application built with a large language model at its core. The typical form is a straight line of input-processing-output: receive user input, the model generates content, return the result. It does not hold execution authority and does not decide the next step in the process.

Story Mapping

  • Assembly arm: Agent, with a closed loop of perception-planning-execution-verification.
  • Response screen: large model applications, a one-way input-processing-output chain.
  • The multi-module structure inside the assembly arm: the Agent's perceiver, planner, executor, and checker.
  • The customer's question "Is it all driven by large models?": the most common rhetorical starting point for conflating two concepts.
  • Adding a microphone to an assembly arm: equating a voice entry point with Agent capability confuses the interaction method with the execution architecture.

In the project

Add a judgment column in solution reviews: does this system end after one answer, or does it need to continue executing, verifying and adjusting after the answer? For the former, a large-model application architecture is enough; introducing an Agent framework adds latency and complexity. For the latter, if it is built only as a Q&A application, it will expose capability gaps in scenarios requiring autonomous decision-making. Drawing the two architecture diagrams separately saves more rework later than a mixed "intelligent system" design proposal.

京ICP备2026014272号