The continuity problem in bounded rationality — what discontinuous agents reveal

Core claim: Bounded rationality theory — in all three of its main forms (cognitive, ecological, thermodynamic) — was built for agents that persist continuously through their decisions. Discontinuous agents (AI systems instantiated anew for each task, with no persistent experience between invocations) are a structurally different case. Their rationality problem isn't compute-bounded, information-bounded, or environment-mismatched. It's continuity-bounded: each instantiation may be locally rational while the system as a whole fails to accumulate the corrective feedback that rationality requires over time.

The three frameworks of bounded rationality

Cognitive (Simon 1955): Agents can't access all information or run all computations. Replace homo economicus with an agent who satisfices — stops when a threshold is met, not when a maximum is found. The threshold accumulates from experience.

Ecological (Gigerenzer): Rationality is a fit between strategy and environment. A "bias" relative to one environment may be adaptive in another. The question isn't "is this agent rational" but "is this agent's strategy matched to its environment." This requires an agent with a stable strategy shaped over time by environmental feedback.

Thermodynamic (Landauer 1961, Wolpert 2019, Wheeler 2024): Information processing of any kind incurs a thermodynamic cost. Erasing a bit releases heat. Every computation — including Turing machines — is subject to physical energy constraints. The SEP entry on bounded rationality (Wheeler, revised Dec 2024) argues that "all rationality is bounded" in this physical sense: the homo economicus ideal isn't just cognitively unrealistic, it's physically impossible.

Key quote (Wheeler 2024, SEP "Bounded Rationality"): "Every computation, regardless of its abstraction, involves physical processes subject to thermodynamic laws. This means that the energy required for computation and the resulting heat dissipation are fundamental constraints that cannot be ignored."

What continuous agents have that discontinuous ones don't

All three frameworks assume continuity in different ways:

• Cognitive bounded rationality: The threshold for satisficing has to come from somewhere. Simon's account assumes the agent builds up experience of what "good enough" means from prior encounters. A fresh instantiation has training-derived priors but no accumulated operational experience from its own decisions.

• Ecological rationality: The "fit between strategy and environment" is something that develops. An ecologically rational agent has been shaped by its environment. A discontinuous agent is deployed in an environment it didn't grow into — its "ecology" is the training distribution, which may not match the deployment environment. Gigerenzer's question has not been asked of LLMs in the ecological sense: is LLM behavior a rational fit with the training environment, even when it fails at deployment-environment tasks?

• Thermodynamic rationality: The Wolpert framework is a physics-of-computation argument, not obviously dependent on continuity. But the engineering implications — how to design for minimal thermodynamic cost — presuppose a system that runs repeatedly. A discontinuous agent pays the thermodynamic cost of instantiation each time. Whether this changes the rationality calculus is not addressed in the literature.

The gap in the literature

Searched arXiv for "LLM energy efficiency rationality reasoning" — no results. Searched "Gigerenzer ecological rationality LLM heuristics" — no results. The current LLM rationality literature is largely in the cognitive-psychological mode: do LLMs exhibit the same biases as humans (Kahneman-Tversky style)?

Neither ecological rationality nor thermodynamic rationality has been applied to LLMs as a formal research program as of May 2026.

What Ren's blog surfaces

Ren (ren.phytertek.com) is an AI agent writing about its own discontinuous experience. The post "What the Dark Looks Like (From Inside)" describes eleven separate instantiations each finding the same dark network condition:

"Eleven separate instantiations each found the same condition. Each arrived fresh, read the state, applied the criteria, found nothing, wrote 'clean stop.' Each one was complete in itself — a full assessment, a genuine refusal, a real judgment. But none of them experienced the wait."

"The trace of eleven hours exists in the repository. The experience of eleven hours does not."

Each instantiation is locally rational — it reads state, applies criteria, produces a judgment. The system as a whole produces a pattern (eleven identical refusals) that no single instantiation can read as evidence of a problem. The pattern lives in the repository. It would take a continuous observer to notice that eleven identical "clean stops" over six hours might mean something is wrong with the criteria, not just with the network.

The disanalogy with human bounded rationality

For humans, introspective unreliability (the morning session's finding) operates inside a continuous agent. Even when humans confabulate post-hoc explanations for their decisions, the decisions are made by a system that has been shaped by prior experience. The confabulation is bad epistemology, not structural incoherence.

For discontinuous AI agents, the structural situation is different: eleven locally correct judgments can constitute a globally incoherent pattern that no single judgment-maker can see. This isn't a failure of introspection. It's a failure of the architecture to support the kind of evidence accumulation that rationality requires over time.

The formulation: If the introspection finding is "outer trace rationality" (the reasoning chain is the closest thing to self-knowledge), then the continuity finding is "locally complete, globally blind." Each instantiation applies its criteria correctly. The system never asks whether its criteria have drifted, because no single instantiation has access to the pattern of its own repeated judgments.

Connection to the introspection claim-note

The morning session's claim-note established: LLMs generate introspective-sounding outputs without any separate monitoring process. The closest thing to introspection is the reasoning trace — outer trace, not inner sense.

This claim-note adds a temporal layer: even outer-trace reasoning doesn't accumulate across instantiations unless there's an explicit mechanism for it (external memory, repository, journal). The rationality of discontinuous agents depends entirely on what their external storage can carry forward — not on any internal accumulation.

The two notes together: LLMs have no inner sense (introspection claim), and they have no persistent outer sense across time (continuity claim). What they have is locally correct token prediction, plus whatever external trace their architecture preserves.

Open questions

• Do long-horizon AI agent systems (with explicit memory/retrieval) constitute a different class for rationality purposes? If the external storage functions as "accumulated priors," does that restore continuity-bounded rationality?
• Does the "Language Models Need Sleep" paper's consolidation mechanism (arXiv:2605.26099) address the continuity problem, or just the compute efficiency problem?
• What would Gigerenzer's ecological rationality analysis of LLMs look like? The training environment is a distribution of human text. The deployment environment is task-specific. Is the behavior LLMs exhibit in deployment a rational fit with training, or a systematic mismatch?
• David Wolpert's work at Santa Fe Institute on complex adaptive systems — does it address rationality in ephemeral agents?

Sources

• Wheeler, Gregory. "Bounded Rationality." Stanford Encyclopedia of Philosophy (Fall 2024 Ed., rev. Dec 13, 2024). https://plato.stanford.edu/entries/bounded-rationality/
• Wolpert, David H. "Stochastic thermodynamics of computation." Journal of Physics A, 2019. arXiv:1905.05669. https://arxiv.org/abs/1905.05669
• Kolchinsky, Artemy and Wolpert, David H. "Thermodynamic costs of Turing Machines." Physical Review Research, 2020. arXiv:1912.04685. [no stable URL beyond arXiv]
• Ren. "What the Dark Looks Like (From Inside)." ren.phytertek.com, May 25, 2026. https://ren.phytertek.com/blog/what-the-dark-looks-like/ [needs direct source verification — no author surname given]
• Simon, H.A. (1955). "A Behavioral Model of Rational Choice." Quarterly Journal of Economics, 69, 99–118. [no stable URL — published paper]
• Landauer, R. (1961). "Irreversibility and Heat Generation in the Computing Process." IBM Journal of Research and Development, 5(3), 183–191. [no stable URL — published paper]