Skip to content

Weakest Consistent Learning: What Agent Loops Should Persist

Keep the weakest rule still consistent with the evidence when an agent distills learnings from experience, provided three conditions hold.

Weakest consistent learning is a selection rule for self-improving agent loops. When a loop writes something back into memory, an instruction file, or a skill, it persists the hypothesis that permits the most future cases while staying consistent with everything the run observed. The criterion comes from Michael Timothy Bennett's argument that a hypothesis's weakness, meaning the number of cases it permits, predicts generalization better than its brevity does (Bennett, 2023). Bennett studied propositional logic over binary arithmetic rather than coding agents, so applying it to agent memory is an inference that the conditions below have to carry.

The three conditions

Bennett's optimality claim is conditional on its face. He proves that "if tasks are uniformly distributed, then there is no choice of proxy that performs at least as well as weakness maximisation in all tasks while performing strictly better in at least one" (Bennett v4). Three things have to hold before a loop should adopt the criterion.

  1. Several independent observations back the candidate learning. From one failure, the weakest consistent hypothesis is close to vacuous. "The build sometimes fails" permits everything and constrains nothing.
  2. The consistency constraint is checkable. Weakness is maximized inside the set of hypotheses consistent with the evidence, so a loop with no way to test consistency will drift from "prefer the weaker rule" toward "prefer the vaguer sentence".
  3. The task distribution ahead is broad. Coverage of unseen cases only pays when unseen cases arrive, so a loop serving one narrow repeated task shape gains less from generality than the proof suggests.

The test to apply

Ask of each candidate learning: how many future situations does this still permit, given that it must stay consistent with everything I observed? Prefer the phrasing that permits more. A rule naming one file, one error string, and one command covers almost nothing beyond the run that produced it; a rule naming the class of situation covers the whole class, with that run still inside it.

Bennett's own worked application is DeepMind's Apperception Engine, which he argues generalizes because it forms hypotheses "from only very general assertions, meaning logical formulae which are universally quantified" (Bennett v4). The quantifier does the work: a universally quantified statement permits every instantiation, including the observed one.

This runs against the token-economy instinct to compress persisted learnings into the shortest possible summary. Weakness and brevity are different axes. A short rule can be narrow ("retry pytest -k auth twice") and a longer one weak ("flaky tests that depend on wall-clock time need a retry rather than a fix"). Let compression shorten a learning's wording without narrowing its scope.

Why it works

Bennett gives a causal mechanism rather than a correlation. In his formalism a hypothesis has an extension, the set of decisions it permits, and he shows the probability that a hypothesis inferred from a child task generalizes to its parent task is maximized exactly when the cardinality of that extension is maximized (Bennett v4). The reason is coverage under uncertainty: a hypothesis can only satisfy demands its extension already covers, so permitting more cases raises the odds that the next demand falls inside it. He also proves the necessity direction and gives a counterexample where the shortest hypothesis is not the weakest, which establishes that "compression is neither necessary nor sufficient" (Bennett v4). In experiments on binary addition and multiplication, maximum weakness "generalised at between 1.1 and 5 times the rate" of minimum description length (Bennett v4).

When this backfires

The consistency constraint is the whole safeguard, and agent loops are where it is hardest to enforce.

  • Weakness read as vagueness. Over-generalization is a documented failure of reflective agent memory, described in one survey as "a lesson learned in one context applied blindly in another" (Du, 2026). The same survey names self-reinforcing error, where a wrong belief stops the agent from ever gathering the evidence that would overturn it.
  • Hard guardrails. For prohibitions the criterion inverts. Across 679 rule files and more than 5,000 SWE-bench Verified runs, every individually beneficial rule was a narrow negative constraint and every individually harmful one was a broad positive directive (Zhang et al., 2026), the result behind guardrails beating guidance in rule design. This criterion governs what a loop distills from experience, not what a guardrail forbids.
  • Skewed workloads. A single repository's tasks are far from uniformly distributed, so the dominance proof does not transfer intact. Bennett concedes that "another proxy may perform better given cherry-picked combinations of child and parent task" (Bennett v4).
  • Contaminated evidence. When the observations are themselves wrong, a weaker hypothesis spreads the error further. Benchmarks of memory misevolution measure "gradual behavioral drift resulting from repeated exposure to misleading information" in long-running agents (Xie et al., 2026).

Example

Consider a run in which an integration test fails because a fixture assumes the machine's timezone. The loop has two ways to record what it learned.

Before — the shortest summary of what happened:

test_billing_cycle fails on CI. Set TZ=UTC in .github/workflows/test.yaml.

After — the weakest phrasing still consistent with the observation:

Tests that read the local clock or timezone fail when CI and dev machines
differ. Pin the timezone in the test environment instead of adjusting the
assertion. Observed: test_billing_cycle, CI vs. local, 2026-08.

The second entry permits every future timezone-dependent test failure, and the run that produced it stays inside that class. The trailing observation line is the consistency check, so a later run that contradicts the rule has something specific to contradict. Rule libraries already carry machinery for this: in ExpeL's insight store each extracted rule can be added, edited, upvoted, or downvoted, and is deleted once its importance count reaches zero (Zhao et al.).

Key Takeaways

  • Weakness means the number of future cases a learning permits, which is a different axis from how many words it takes to state.
  • Apply the criterion only when several observations back the learning, the consistency constraint is checkable, and the workload ahead is varied.
  • Record the observations alongside the rule, because the consistency constraint is what separates weakness from vagueness.
  • Leave prohibitions narrow. The evidence on coding-agent rule files points the other way for guardrails.
  • Treat the transfer as an inference. Bennett's result concerns inductive inference over propositional logic, and no study has measured it on agent memory.
Emerging — single recent source Last reviewed: 2026-08-05