CHAPTER 07 / 7 MIN READ
Keep becoming
Separate the thing being edited from the time at which it is edited.
START HERE
An agent has been working for hours. It discovers a better way to navigate its environment. Must it finish, reset, and recompile before using that improvement?
What can change and when it can change are two independent design questions.
1.The timing axis changes the story.
An offline optimisation phase can evaluate completed attempts and install a new configuration for the next run. Reset-free refinement instead changes relevant state while the environment continues.
Continual Harness refines prompts, sub-agent specifications, skills, and memory from trajectory information during a run. Its separate co-learning experiments also include model updates. These are distinct interventions, even when they appear in the same paper.
2.A session needs more than a transcript.
Prime Agent combines a persistent programming environment, recursive subagents, and retained harness state with execution and recovery machinery. Think of the ongoing task as more than a single input-output exchange.
For a long task, the system must know which sessions exist, what has happened, what state survives interruption, and what resources were spent. Losing that bookkeeping can make an otherwise capable model fail.
3.Persistence makes changes consequential.
A local correction can disappear when the run ends. A persisted skill or instruction can shape later behaviour. That is the benefit of accumulating experience—and the reason a bad refinement deserves special attention.
Do not collapse persistent memory into weight learning. A file edit changes external state. A weight update changes the model’s parameters. The public atlas preserves this distinction even where the inherited ladder is ambiguous.
The ongoing factory
BEFORE
An improved production procedure is available only after the next offline optimisation cycle.
AFTER
The procedure is revised during the same continuing task and becomes available at a later turn boundary.
Timing affects when an improvement can help. It also affects how quickly an unchecked error can spread.
PAUSE & PREDICT
An agent saves a new skill file during a run. What changed?
The corpus’s highest rung includes Prime Agent by inheritance. That placement is not proof of a demonstrated online weight-updating system in Prime Agent.
GO TO THE SOURCE
Continual Harness2026 ↗Prime Agent2026 ↗This chapter is an original teaching synthesis. Open each research note for source links, findings, and classification caveats.
TAKE IT INTO YOUR OWN WORK
For a workflow of your own, specify what can change during a run, what must wait for evaluation, and what should survive a restart.