article research
Continual harness: from paper to Prime Agent
The Continual Harness paper formalizes reset-free online adaptation over prompts, subagents, skills, and memory in embodied domains. Prime Agent adapts the abstraction to a coding/research harness through /refine, but the paper's model-training loop and Pokemon results are not a direct product evaluation. Capability dependence and scope ambiguity remain.
First-party announcement, documentation, release, or installer.
Continual Harness paper
The Continual Harness paper studies reset-free online adaptation over prompts, subagents, skills, and memory in embodied Pokémon environments. The harness changes during experience instead of resetting to the same fixed configuration for every episode.
Its reported experiments concern that research setting, not a direct benchmark of Prime Agent on coding or research tasks.
CRUD surface
The paper frames adaptation as CRUD operations over mutable harness components: instructions and supporting structures can be created, inspected, changed, or removed as new trajectories provide evidence.
This is harness-state adaptation rather than a claim that the model weights themselves are retrained online. The quality of the update still depends on the objective and the model's ability to infer a useful change.
Prime Agent mapping
Prime Agent maps this abstraction to /refine, which proposes small changes to prompt notes, memories, skill descriptions, or subagent specifications and records reversible refinement IDs.
The mapping is conceptual and operational, not an identity between the paper's training loop and the product. Prime Agent applies the idea to a coding and research harness with its own turn boundaries and persistence stores.
Capability dependence
The paper reports high variance and a capability floor: not every underlying model benefits equally from continual harness adaptation. A model must be able to diagnose trajectories and propose useful state changes before the loop can improve behavior.
Consequently, the existence of /refine does not predict the same gain across models or tasks. It can also reinforce a poor objective, as the cited Factorio warning illustrates.
Scope ambiguity
Prime Agent's exact persistence scope for every refinable artifact is not fully resolved across releases. The reviewed sources establish a mutable surface, but not that each edit applies to every project, session, or future runtime.
Verify where a proposed change is stored and whether it survives the lifecycle that matters. Do not infer a rewrite of the immutable base system prompt from a successful harness refinement.