Concepts/Concept
Control Plane and Narrative Plane: The Dualism of Execution-Type Agents
From the Dongfang Yiteng execution agent (case by Bo Liang)
Definition
The runtime context of an execution-type agent should be split, from the very first step, into two planes of different character.
The control plane is mechanical, deterministic, and discrete: a small set of signals that decide which branch the program takes next — small talk ends the turn, a query triggers retrieval, a task triggers planning. The narrative plane is semantic context in natural language: it feeds every reasoning step and, after reasoning, produces new control signals. Both run in parallel. Without control signals the engineering has nothing to drive it; without narrative context each reasoning step has no raw material.
This dualism is the cognitive foundation of the whole system. It answers the question that stalls most people new to agent work: where do you even grip something described with words like "intelligence" and "uncertainty"? The move is to convert the live and uncertain into the dead and controllable first — because only the dead can be controlled, and control is mechanical by nature.
How it works
The control plane must be mechanized. The narrative plane has its own management — semantic compaction, assembly, writing, scenario-based injection — but that is not strict state control; it is closer to steering. Bo Liang puts the difference in terms of riding a horse: the control plane is the reins and stirrups, which keep the horse going where you intend most of the time; the narrative plane is the horse's mind, which you can steer but cannot guarantee will obey the next instant. Kubernetes splits a cluster into a data plane and a control plane on the same judgment.
Skip the split and you pay for it. If the execution order of a business process is handed to the model as narrative context, in the hope that it runs the right steps inside a loop, the observed result is that it sometimes runs correctly and sometimes skips or repeats a step — and when it fails is neither controllable nor predictable. The root cause is that an LLM is a probabilistic model; handing deterministic control flow to it for on-the-spot judgment conflicts at the root.
Dongfang Yiteng first saw these two planes at the earliest intent-recognition step: a single capability's output already separated into a control signal that drives where the program goes, and narrative content for downstream reasoning to consume. Every architectural trade-off after that was built on this line.
When you need it
When your agent executes a chain of business steps with strict state dependencies, rather than generating content. Splitting the control plane's state further is the unified session state and its separation of powers; how the narrative plane organizes itself is anchor, ledger, collection.
This is an ADPS blue-book concept. Back to Concepts or the pattern matrix.