Agentic control

Most language-model agents act one step at a time: observe, choose an action or a tool call, observe again. The model sits inside the control loop, so every step waits for it, and every step costs a call.

Agentic control takes the model out of that loop. The model writes the controller instead: programs that perceive, act and check their own results, hand-offs to learned policies, and reflexes that run on every control step. The controller runs on its own; the model is asked again only at decision points, when a program ends, when something escapes every layer below it, or when what comes next depends on what has not been seen yet. An episode then takes a handful of model calls, and nothing below the model ever waits for one.

Four layers, four speeds

Layer	Written by	Runs	What it does
Language model	any model you configure	at decision points, in seconds	plans, writes the next program, reads its result
Programs and learned policies	the model (programs); training (policies)	per skill call	compose skills and tools; hand a whole task to a learned policy where code falls short
Reactive layer	the model	every control step	guards and controllers that hold, stop or correct, and raise events
Safety layer	the harness, fixed	every command	limits nothing learns or rewrites, whoever wrote the command

Each layer handles what it can at its own speed and passes up only what it cannot: a guard that trips raises an event to the program, a program that fails ends with its error, and only then does the model look. The episode runtime has the details of each layer; Programs and decision points explains how the model decides what to write and when to hand a task to a learned policy.

Code where code works, policies where it does not

A program is good at what can be written down: find the object a name or a label identifies, move through free space, check a result and try again. A learned policy is good at what cannot: skilled contact many times a second, from what it was trained on. At the start of a task the model reads each policy’s description and decides who does the task. A task a policy was trained on goes to it whole; the programs keep what the policy cannot do.

Anything that acts

The harness meets an environment through observations and commands only. The same agent runs on a benchmark’s official environment, installed on your machine, or on a task from The Simulator; the language model, the learned policies and the tools are each one line of configuration. Skills and Services and environments describe how capabilities plug in.

Results that name what produced them

Every run records a lock and an identity: the prompt, the skills, the models and every setting that decides what it computes. Where the models run is not part of the identity, so the same agent on a larger or smaller machine is the same agent. Configuration and identity.