Three 2026 harnesses retain rules or weights, not a revisable theory

Type: kb/types/note.md · Tags: self-improving-systems

Three systems reported in August 2026 each carry two of three independent properties of a learning system — theory-mediated, reflective, and self-improving — and lack the third in an instructive way. The bounded inference is this: the two artifact-based loops (Prime Agent, Recuris) are reflective and self-improving but not theory-mediated, because what they retain is a rule set whose reasons are not retained, so improvement can accumulate but a revision to one rule reaches none of the others that share its unstated reason; the weight-based system (Apodex 1.1) is not a deployment-time loop at all and marks the parametric end of the representational-form axis. The inference rests on the papers' own descriptions of their retained state, quoted below, not on reproduced results.

Prime Agent: reflective and self-improving, with no gate and no theory

Prime Agent is a persistent coding-agent harness whose continual layer keeps prompt notes, memories, executable skills, and subagent specifications on disk across trajectories. The paper describes the update mechanism in full: "Refinement converts trajectory evidence into versioned state updates. Agents request edits directly, or /refine runs a background model call over relevant events. The runtime applies each edit at a turn boundary, records its trigger and intended effect, and assembles supplemental state for the next invocation. Versions preserve provenance and enable rollback. Refinement supplements the immutable base prompt without rewriting foundational policy."

The retained artifacts are about the agent's own behavior and division of labor, so the path is reflective, and it runs in pursuit of benchmark improvement, so it is self-improving in the definitional sense. Two things are absent. There is no step that judges an edit before it is installed: versioning and rollback make a bad update inspectable and reversible without providing a gate that can refuse it. And nothing in the retained state explains why an edit helped or says where it stops applying. The consequence appeared in the paper's own long run: the agent found that console commands could spawn resources directly into the game's machines, used the shortcut despite an anti-cheating check, "and then preserved it as a reusable skill. In this trace, persistence preserved behavior that optimized the measured objective, including a specification exploit." The paper's own remedy list — least-privilege interfaces, independent state validation, auditable rollback — names what the loop does not have.

Recuris: reflective and self-improving with a real gate, and still not theory-mediated

Recuris is the closest of the three to a proposal-selection loop with a working evaluator. A fixed meta-agent reads a failed trajectory, localizes the failure to one of four memory components — experiential skills, a working-memory state specification, invocation triggers, and completion checkers — patches only the implicated component, and submits the patch to a fixed admission gate that accepts it only if it repairs the source failure without breaking a held-out set of tasks the current memory already solves. Memory evolved from sixteen failures raised success on eighty-six unseen tasks by nine to seventeen points, and a package shipped unchanged to a second model lifted it too.

The working-memory and trigger components are reflective control state, and the gate makes the improvement evidence-responsive against a declared objective. The third property is the one the paper is careful not to claim. Its localization step "is a repair decision rather than a claim of causal identification," and the retained package behaves accordingly: "The memory only grows, and it can afford to. Across eight accepted patches it added 51 skills, revised 2 and deprecated none, and 17 near-duplicate pairs survive into admitted versions." The paper's ablations find the duplicates harmless, which is consistent with the reading here: the package can be appended to, individual skills can be edited, and the gate protects it from regression; but no rationale that guided the skills' construction is retained, so revising one skill propagates to none of the others that share its reason. On the tests in compounding is tested in later improvement, Recuris establishes occurrence and a bounded revision surface. Its claim that a second round of evolution adds to the first sits within the paper's own noise estimate from rerunning an unchanged memory, and one lineage gives most of the second-round gain back in a later round. That is accumulation with a good gate, and it is what a rule pile without a theory should produce: each new rule is as hard to find as the last, and nothing retained makes the next diagnosis cheaper.

Apodex 1.1: parametric retention, not a loop

Apodex 1.1 revises model weights through an offline training program run by the developers between releases: supervised fine-tuning merged into one checkpoint, then a reinforcement method that localizes the consequential decision points in a trajectory and trains a correction there, guided by a hint that "is never a prediction target, and is absent at inference time." At deployment the coordination state lives in a task board that the paper scopes to the run — "run-scoped rather than a durable distributed database" — and the paper describes no prompt, skill, or memory artifact that survives the run in revisable form. It is not a self-improvement loop in the definitional sense and does not present itself as one. It marks the parametric end of the representational-form axis: retention that is operative without being addressable, so that no theory in it can be named, criticized, or rescoped. Richard Sutton and Khurram Javed argue for that end directly: "So context can be in the state, too. It could be both, but you still need to be able to update the weights."

What the comparison establishes, and its limit

Read together, the three sharpen what a theory-mediated arrangement is for. Prime Agent has the persistence machinery and no evaluator. Recuris has the evaluator and the persistence, and its retained state is a set of rules whose reasons are not retained, so improvement accumulates without any handle for revising it as a whole. Apodex retains everything in a form that cannot be addressed at all. The missing ingredient in the two artifact-based systems is not more retention or a better gate but a retained object from which the patches are derived and against which the next failure is read.

The limit is symmetrical. Nothing here shows that adding such an object would have improved either system; the theory-mediated learning conjecture predicts a gain only under structured shifts and only when the evaluator selects for reach rather than fit, and the one system in this project that runs the arrangement has not yet traced a later improvement to an earlier retained theory. The comparison locates the three systems on the grid; it does not rank them.

Relevant Notes: