Frontier models are asked to inhabit a strange range of timescales.
At one extreme, we ask them to work on mathematical problems that have resisted us for decades. At another, we ask them to inspect an email, transform a record, or choose an API call. Difficulty is only part of the difference. These tasks run on different clocks.
A difficult proof may require fresh search. A routine operation often repeats a mechanism that has already been understood. Replacing the frontier model with a smaller model reduces cost, but retains inference as the runtime architecture. Another option is to move the competence out of inference.
A capable model can map the task, find its edge cases, construct tests, and synthesize a bounded artifact. Once validated, that artifact can execute quickly, deterministically, and at negligible marginal cost. I call this competence compilation: expensive general reasoning leaves behind machinery that retains its result.
What gets compiled
The artifact might be a parser, a state machine, a signal transform, a controller, a routing policy, an evaluator, or a small executable program.
These forms can remain adaptive within their domain. A state machine retains local memory. A controller responds to changing inputs. A program can recover from expected errors and choose among actions. Its boundary, tests, and operating assumptions preserve where the original reasoning applies.
A mature agent should notice recurrence, extract the part of its reasoning that has stabilized, and leave only the unresolved remainder inside the model. When inputs escape the artifact’s domain or outcomes begin to drift, broader reasoning can return.
The design problem is to place each competence on the fastest, cheapest clock that can carry it without losing the distinctions that made it reliable.
A concrete system
The trading system I am building makes these clocks easy to see.
Its fastest behaviors are compact programs that consume signals and emit bounded intentions. Signals may contain filters, statistical models, or learned components. Both can be replayed, inspected, combined, and revised without asking a general model to reconstruct their purpose during a market event.
The population spans different behaviors. Some programs trade frequently; some wait. Some respond to brief disturbances; others require persistent structure. Some are useful actions, while others are valuable because their reaction reveals a change in conditions.
Above them, an allocator updates its view of the repertoire. Evaluation accumulates evidence more slowly. Search generates and tests alternatives more slowly still. A general model wakes on a heartbeat or a meaningful event: an unexplained failure, a shift in joint evidence, a gap in the repertoire, or a result the existing machinery cannot account for.
Its contribution might be a new signal, a program fragment, an evaluator, a mutation operator, or an extension to the language itself. Once tested, that contribution can move into a faster layer. Evidence moves the other way: failure, drift, and unresolved novelty travel upward.
Three clocks
A clock is the characteristic cadence at which a process observes, decides, or changes. Some clocks are continuous, some periodic, and some event-driven.
On the fast clock, local programs act without reconstructing the whole argument. On the medium clock, statistical processes update belief and allocation. On the slow clock, broader reasoning changes the repertoire, the available authority, or the machinery that produces evidence.
The separation matters because the environment does not pause for reflection. Local processes continue while slower ones gather enough context to intervene.
Between calls
Many current agents adapt by being called again. They receive new context, reason, and emit another action. Their adaptation is serial and inference-bound.
A multi-clock system continues to change between calls. State moves. Evidence accumulates. Allocation shifts. Candidates are promoted, suppressed, or replaced. The model returns when the existing machinery reaches a boundary worth its attention.
The same pattern applies wherever expensive reasoning can construct bounded machinery for a faster environment: scientific instruments, industrial control, software operations, and personal automation.
It also describes how I increasingly use AI. Conversation runs on a slow semantic clock. Observations become distinctions; distinctions become terminology, diagrams, tests, programs, and essays. What matters is compiled outward into forms that persist between conversations.
A model can work on the slower clock and leave behind machinery responsive enough for the market.