← Back to Concept Contrast

Concept Contrast Fable

Authority boundaries and autonomy levels

Conclusion first:The authority boundary defines "what can be done"—which operations are permitted, which data is accessible, which systems can be reached. The autonomy level describes "how much it is allowed to decide on its own"—within the boundary, whether it can act directly without asking when situations arise.

The authority boundary defines "what can be done"—which operations are permitted, which data is accessible, which systems can be reached. The autonomy level describes "how much it is allowed to decide on its own"—within the boundary, whether it can act directly without asking when situations arise.

Fable

Fences and keys

A large workshop opened on the outskirts of town. The owner hired a group of craftsmen and assigned each person a work area. Zhang San's work area was the woodworking room, with saws, planes, sandpaper and a workbench. The owner pointed at the fence and said: you can use all the tools and materials inside this fence; outside the fence — the welding room, the paint room, the finance office — you cannot touch. This is the workshop's scope of authority.

Zhang San worked every day in the woodworking room, but every time he changed tools he had to run to the boss's office for approval: can I use saw No. 3 for the next board? The sandpaper is used up, can I get more from the warehouse? The customer wants the table legs shortened by two centimeters, can I change them? The boss was fed up, and Zhang San was exhausted too.

The foreman said the key thing: the fence controlled what Zhang San "could touch," but it did not answer which things inside the fence he could decide on his own and which required approval. Sawing planks and swapping in the same model of sandpaper did not require asking the boss. Changing design dimensions, switching to more expensive wood, or pushing back a delivery date required approval. The foreman suggested giving each craftsman an "autonomy key": a copper key for daily decisions, a silver key for scheduling decisions, and a gold key for company-wide decisions. Zhang San received a copper one; the foreman held the silver and gold ones.

Once the rules were running, efficiency rose sharply, and Zhang San no longer had to run to the office over "which saw number to use." But a new problem arose: once Zhang San received temporary authorization for the silver key, he used it to reorder the priorities of three customer orders—he got the deadline wrong on one urgent order, and the customer complained. The retrospective found that the key defined the "level" of autonomy, but did not define the "safety confirmation points" within each level—which reordering actions require secondary confirmation.

The workshop changed the key system into a table format: each level listed specific operation checklists and confirmation rules. Bronze level: fully autonomous. Silver level: sixty percent autonomous, forty percent requiring sign-off. Gold level: all requiring two-person co-signature. Keys were no longer vague levels, but passes with detailed permissions.

One year after the workshop began operating, the boss shared at an industry conference: authority boundaries govern the fence—from the outside, who is in which enclosure. Autonomy levels govern the keys—from the inside, who within the fence can decide for themselves. Both are needed, and the granularity of the keys is much finer than that of the fence.

Later the workshop drew the fence and the keys on the same site board: red fence lines marked authority boundaries, and yellow keyholes marked autonomy levels. Which craftsman was in which zone and holding which key could be seen at a glance by the guard on duty.

Later a new rule was added: before issuing a key, check the nameplate—do not hand a silver key to someone inside a copper fence. If the key does not match the fence, no amount of skill will help.

Authority Boundary

Defines the scope an agent can access and operate—including systems, data, APIs, and physical devices. It answers "what you can and cannot act on," and is usually enforced by permission configuration and security policies.

Autonomy Level

Within the boundaries of authority, how much an agent can decide autonomously without human approval. It answers whether, when a situation arises, it can act directly or must first ask for instructions, and is usually divided into multiple levels, from fully automatic to requiring human confirmation.

Story Mapping

  • Workspace fence: authority boundaries that define the accessible resources and scope of operations.
  • Copper/silver/gold keys: autonomy levels, different degrees of autonomy within the same authority boundary.
  • Zhang San asks his boss for a saw: the authority boundary is drawn, but the autonomy level is too low.
  • Zhang San used the silver key and got the urgent order date wrong: the autonomy level was raised, but an operational safety confirmation point was missing.
  • Tabular key (each level has an operation checklist and confirmation rules): a refined autonomous control strategy.

In the project

When designing an Agent system, draw the fence first, then assign the keys. At the fence level: list a permissions inventory that clearly states which APIs the Agent can access, which database tables it can read and write, and which file directories it can operate on. At the key level: grade operations by business risk. Low-risk operations (queries, summaries, format conversion) get the highest autonomy level, and the Agent handles them on its own. Medium-risk operations (modifying data, sending messages, canceling orders) require a confirmation step. High-risk operations (deleting data, changing permissions, moving funds) require mandatory human approval. The worst mistake is drawing a very large fence but giving everything the highest autonomy level—that is equivalent to having no safety mechanism.

京ICP备2026014272号