← Back to Concept Contrast

Concept Contrast Fable

Agents and AI assistants

Conclusion first:Agents and AI assistants are both called "AI that does things for people," but their positions in the workflow are completely different. Once an agent receives a goal, it finds its own path, executes on its own, and corrects its own errors—you only see the result. An AI assistant offers suggestions, looks up information, and checks errors while you work—you always keep your hands on the wheel. Upgrading an assistant to an agent is not a matter of adding one permission; it is a redesign of the entire chain of responsibility.

Agents and AI assistants are both called "AI that does things for people," but their positions in the workflow are completely different. Once an agent receives a goal, it finds its own path, executes on its own, and corrects its own errors—you only see the result. An AI assistant offers suggestions, looks up information, and checks errors while you work—you always keep your hands on the wheel. Upgrading an assistant to an agent is not a matter of adding one permission; it is a redesign of the entire chain of responsibility.

Fable

The ghostwriter and the bookkeeping assistant

On the old street there were two people who helped others get things done. Mr. Fang was a proxy agent: the client gave him the goal, boundaries, and seal, and he ran to the government office himself, supplemented materials, queued for stamps, and finally delivered the completed documents back. Awen was an accounting assistant: when the manager kept the books, he sat beside him flipping through old ledgers, reminding him of due payments, and checking the numbers.

Both save people effort, but what the employer hands over is different. Hiring Mr. Fang means entrusting an entire process; hiring Awen means making fewer oversights and miscalculations while doing the work, with the steering wheel still in the shopkeeper's hands.

One year, Manager Wang felt that Awen knew the business well and allowed him to directly correct obvious erroneous entries. In the first month, Awen corrected three entries and all were right; in the second month, he flattened a special write-off, and when an old client came to reconcile, they found that the original agreement had been erased.

Shopkeeper Wang first thought of giving Awen more ledgers so he could learn more comprehensively. But exceptions in business changed daily, and much context appeared only at the negotiation table. No matter how diligent the assistant was, he did not know which sentence overrode which old rule.

So Awen stepped back to a supporting position: flag anomalies in red, give sources, and wait for the shopkeeper's nod before acting. Mr. Fang, by contrast, had a letter of authorization from day one, clearly stating what could be handled on his behalf, which amounts required checking back, and who signed in the end.

Later, peers asked Boss Wang why he did not also turn Awen into Mr. Fang. He said the difference between the person beside you giving reminders and the person who takes over the whole matter is not intelligence, but scope of authorization, responsibility for action, and exception handling.

In the end, Old Street did not confuse the two kinds of helpers. Giving an assistant too large a hand lets it reach past its master; giving an agent a missing boundary makes it stop at the door. What truly must be decided first is which piece of work the person is prepared to hand over.

Later, Manager Wang added new business and no longer first asked "who is more capable," but instead cut the work apart: which steps merely reminded him to look, which steps could be entrusted in full, and which exceptions had to come back for a signature. After clarifying these, both Awen and Mr. Fang were easier to use, and the shop no longer mistook saving effort for delegating authority. This judgment later saved him a lot of money. When he needed assistance, he did not buy a whole set of proxy services; when he needed results, he did not expect the person reminding him nearby to secretly take responsibility. Both types of helpers can create value, provided the work boundary is clarified first. This is more useful than arguing about names, because what customers pay for is the location of responsibility.

Agent

An AI system with autonomous perception, decision-making, and execution capabilities. After receiving a goal, it plans its own path, calls tools, handles exceptions, and outputs completed results rather than suggestions.

AI assistant

An AI system aimed at enhancing human work efficiency, providing assistance such as information retrieval, content suggestions, and error checking, while decision-making and execution authority always remain in human hands.

Story Mapping

  • Ghostwriter (Agent): completes the entire task independently; the client retains only signing authority
  • Bookkeeping assistant (AI assistant): flips through old accounts, reminds, and cross-checks beside you, but never makes decisions for the boss
  • The lesson from Manager Wang granting Awen permissions: upgrading an assistant is not the same as rebuilding an agent; without the owner's unique information, errors will occur
  • Awen returned to the assistant role: after permissions were revoked, a "mark in red but do not modify" reminder mechanism was added
  • The boundary the master drew on day one: an agent's authorization boundary is designed, not something that naturally grows out of iteration

Applying it to Andy Ren Xu's project

Before deciding to build an agent rather than an assistant, first draw a decision tree: which judgments the model makes on its own, which must wait for human confirmation, and how to degrade when confirmation times out. If you cannot draw this tree, it means your "agent" will quietly degenerate at key nodes into a half-finished product waiting for a human to take over. Conversely, if what the customer wants is just a tool that helps people be more efficient, do not force in an autonomously acting solution—the authorization design and exception-handling costs you save are enough for two rounds of user experience optimization.

京ICP备2026014272号