← Back to Concept Contrast

Concept Contrast Fable

Agents that make decisions and autonomous driving that guards boundaries

Conclusion first:An agent makes decisions in an open domain—when facing new tools, new goals and new contexts, it can reorganize its behavior. Autonomous driving stays within a bounded domain—such as driving, reasoning, form-filling—and achieves extremely high reliability within the boundary. Treating autonomous driving as a scalable agent will lead to loss of control beyond the boundary after launch.

An agent makes decisions in an open domain—when facing new tools, new goals and new contexts, it can reorganize its behavior. Autonomous driving stays within a bounded domain—such as driving, reasoning, form-filling—and achieves extremely high reliability within the boundary. Treating autonomous driving as a scalable agent will lead to loss of control beyond the boundary after launch.

Fable

The automatic boat at the canal lock and the pilot who runs the open sea

There were two kinds of boat operators in town. On the canal ran a row of identical iron-hulled boats, loaded with programs in their holds, slowing down on their own at the lock, queuing, and passing once the water level equalized. On the open sea ran Old Gu, who for over thirty years relied only on a star chart and a bamboo pole. Out-of-town shipowners came to town to hire crews for cargo, thinking the two kinds of boats were the same — both could carry goods from south to north.

In March the water rose, the canal lock malfunctioned, and the water level no longer followed procedure. The iron-hulled boat swayed back and forth in place—the water level conditions in the program could never match, and no one could rewrite it. Old Gu happened to be on the inland river that day; one look at the water surface told him the upstream floodgate had been opened, so he took a detour of seventeen li along a tributary and delivered the goods that night. The townspeople said Old Gu had won.

But in peak shipping season, the canal had to make two hundred trips a day. Even if Old Gu had three heads and six arms, he could only make four trips a day, and every detour, judgment and decision took effort. The iron-hulled boats queued smoothly, moved at consistent speed, and measured fuel consumption precisely to the decimal. In the peak-season total accounts, the iron-hulled boats' shipping volume was more than twenty times Old Gu's. The mayor summed it up at the dock: one can handle surprises, but there is only one of him; the other cannot handle surprises, but can be replicated infinitely.

Later, someone came to Old Gu to learn how to sail the open sea. Old Gu compiled a lifetime of sea conditions into a booklet. His apprentice memorized it for a year, yet still capsized once at sea—the waves that day were not in the booklet. The iron-hulled ship's engineers wanted to extend the canal program to the open sea. After ten months of writing, they discovered that the number of tidal variation combinations alone exceeded the branches the program could handle.

Broken down, Old Gu's advantage lies not in any single route, but in his ability to judge what situations he has never seen and which direction to try next. The iron-hulled ship's advantage lies in all the situations it has seen; it clears a lock three times faster than Old Gu. What lies outside the boundary is called surprise; what lies inside the boundary is called throughput.

The town set the rules. When canal capacity was tight, the iron-hulled boat handled daily volume, while Old Gu handled urgent items and detour orders. All sea transport was handed to Old Gu—if the iron-hulled boat did not capsize, fine, but if it did, the entire lock would be blocked. One boat traded capacity for the ability to handle surprises; the other traded the ability to handle surprises for capacity.

A mistake new boat owners often make: thinking that upgrading the iron-hulled boat's program will let it sail the open sea. They upgraded the water-level judgment, added a wave-height module, and embedded wind forecasting, but the first time three anomalies occurred at the same time, the boat spun in place. Old Gu passed by, took one look and said: You're not teaching it how to sail the sea; you're drawing every sea route into the blueprint of a canal lock gate.

Agent

An entity that autonomously perceives, reasons and acts in open domains, able to reorganize behavior for new goals and new contexts, suited to varied and complex task environments.

Autonomous driving

An automated system that completes specific tasks within a bounded domain, with clear boundaries, reproducibility, and optimization to extremely high reliability, but uncontrollable behavior beyond those boundaries.

Story Mapping

  • Iron-hulled boats on the canal: autonomous driving, replicable within a bounded domain, high throughput, low fault tolerance
  • Old Gu goes to the open sea: agents rely on judgment and adaptation in open domains, with low capacity and high flexibility
  • Floodwaters detour seventeen li: the agent's advantage when facing unexpected behavioral reorganization
  • Peak-season throughput differs twentyfold: the scaling advantage of autonomous driving within a limited domain
  • An attempt to write an open-sea program: expanding a bounded-domain solution to open domain causes a branch explosion
  • New interval locks in division of labor: pricing by boundary theory—use autonomous driving within the boundary, agents outside it

Down to task automation selection

First, draw the task boundary clearly. Within the boundary, if all possible states are enumerable, failure modes are known, and throughput demand is high—choose an autonomous driving approach; replicability is worth more than flexibility. If the boundary is fuzzy, states are not enumerable, and failures require on-site judgment—choose an agent; flexibility over reliability. The most expensive mistake is forcing an autonomous driving approach beyond its boundary: adding one rule module at a time, until the system spins in place when the boundary fails. The signal is: if your exception-handling code exceeds thirty percent of total code volume, this task is not in a bounded domain.

京ICP备2026014272号