Rather than attempt to maintain an entire task within a single large model, we derive behavior from small, bounded loops that retain only the context they require. These loops assemble to form workflows that remain stable in the face of increasing complexity.
The usual method for using a language model to complete a task is to present it with the entire job at once. A single long prompt conveys all aspects of the task: roles, rules, history, boundaries of safety and immediate requirements within a single context. The model is expected to maintain consistency with all elements at each stage of processing. This approach is effective until the limits of the context are reached. Then difficulties appear and do so very gradually.
The attentional system is flat, so conditions that should control behavior compete with normal information processing rather than controlling it. As a conversation unfolds, the amount of normal information processing grows rapidly and progressively displaces conditions of control. Behavior drifts, not because the model was argued out of its instructions, but because those instructions were progressively outweighed by normal information processing.
This note is about the unit we build with to keep that from happening.
What an atomic loop is
An atomic behavioral loop is the smallest self-contained element of useful behavior: a brief sequence of perceiving, deciding, acting and verifying appropriate to the performance of a single task. It is literally atomic, performs a single function, contains only the information required for that function and can be analyzed, tested and trusted independently of other functions. It is a loop, not merely a generator of responses that terminate abruptly, but a process that integrates an existing situation, generates a judgment, produces an action and verifies the result against the conditions it was given before anything is allowed to proceed.
The important change is in the treatment of context. In the monolithic approach, context is part of the conversation and constitutes a continuously expanding window that must be shared by all steps. In the atomic approach, context is part of the loop. Each loop receives a tightly controlled and highly specific portion of context: the task to be performed, the constraints that apply, and nothing else. The conditions that should hold thus accompany the unit of work and remain close to it, where they cannot be diluted by information of no relevance to the task at hand.
The conditions that govern a behaviour should travel with it, not compete with everything else in one shared window.
Why small and bounded beats large and open
This is not a matter of preference for tidiness. It reflects the same requirement for adherence at the level of system design. A constraint imposed within a small, bounded context is one that cannot be overridden, because there is no extraneous information within the window to override it.
Bounded context is not a technique for fitting more into the model. It is an alignment primitive. It represents the means by which a value, a boundary of responsibility or a requirement for safety remains effective and does not dissipate with increasing length of exchange.
It also alters the locus of failure. In a long context, an initial small drift propagates silently throughout all subsequent work and is apparent only at a point where its origin is no longer identifiable. When work is decomposed into atomic loops with separate checks, errors are detected at the point of origin and prevent propagation. Failure of a loop to satisfy its own check does not result in transmission of the error to subsequent work. At a sufficiently high level of complexity, this is of major importance: A series of limited, controlled operations maintains its accuracy of operation, whereas a single uncontrolled operation leads progressively to loss of accuracy.
Modelled on how people actually work
The rationale for designing in this way is not solely architectural but reflects the ways in which skilled professionals operate. An experienced clinician does not attempt to retain all aspects of a case in an undifferentiated whole; instead, the physician carries out brief, highly practiced routines of risk assessment, comprehension, a return to the limits of their professional role and appropriate timing for each. Similarly, an experienced sales representative performs a series of qualified, exploratory, structured and confirmatory operations within limited boundaries rather than as a continuous process of improvisation.
Ultimately, close examination of expertise reveals that it is largely a capacity for execution of brief, highly reliable loops and for appropriate judgment concerning the timing and sequence of their application.
This is why our work is interdisciplinary, and not purely technical. To represent the behavior of an expert in a domain, it is necessary to represent the way that expert actually thinks, directs attention, resolves uncertainty, and operates at the limits of safe practice. This requires substantial input from psychology and mental health as well as from philosophy and engineering.
We are not attempting to replicate the operation of a machine predicting the next output token. We are developing a cognitive representation of the performance of a skilled, careful human operator, and representing that performance as behavior that can be executed with full fidelity by a system.
How loops compose into adherence
Atomic loops are not the entire system, but are what the entire system is constructed of. They compose. A workflow results in a sequence and organization of loops that are all bounded and validated, with conditions of control reimposed at each transition from one loop to the next rather than lost to degradation over a single long pass. When a segment of work is heavy, it is subdivided rather than overstuffed, to avoid requiring any single step to maintain greater fidelity than it can possibly achieve.
Adherence is not a quality we expect to maintain for an ever-lengthening conversation. It is a quality that we construct at each loop of the process from units sufficiently small to remain faithful and yet sufficiently structured to allow integration. This same strategy applies equally well to a commercial pipeline in which the cost of drift is a wasted account or an off-brand message, and to a highly sensitive situation in which the cost of drift represents a crossed boundary of critical importance.
Consequently, for all applications, it is essential to maintain small units of behavior within well-defined boundaries of context. Always test the unit before using it as the basis for subsequent processes.
It's an unglamorous notion, and we believe it is a correct one. Reliable behavior at scale is not a single large mind that maintains complete awareness at all times. It is the result of many small loops that remain faithful to their particular conditions and are carefully integrated to produce a stable whole.