Nature Perspective: autonomous agents need internal states they must keep in range

Factorization is the word doing the work ^[4], and it is an architecture commitment rather than a loss-function tweak. Inside a single undifferentiated state vector, a falling battery reading is one more observation, worth whatever the external reward happens to correlate it with. Split the vector, give the internal variables their own dynamics and set-points, and a departure from a set-point becomes an error signal that exists whether or not the external task is legible at that moment. The abstract makes precisely that claim for internal states: universally available, intrinsically valuable, stable references that modulate learning and behaviour while the outside environment changes ^[5].

The lineage is declared rather than implied. Ashby's *Design for a Brain* (1960) sits in the reference list ^[9], next to Friston and colleagues on the free energy principle from 2023 ^[10] and Man and Damasio on homeostasis in soft robotics ^[11]. The authors describe the piece as a Perspective integrating cybernetics with theories of life, reinforcement learning and neuroscience ^[7].

The gap being claimed is narrower than the framing suggests, and the citations show why. "Human-level control through deep reinforcement learning" (2015) and "Language models are few-shot learners" (2020) appear as the advances whose limits are named ^[12][1], but so do a 2022 survey of intrinsically motivated goal-conditioned agents and a 2022 review of continual reinforcement learning ^[13]. Intrinsic motivation already puts a signal inside the agent. What is different here, on the abstract's own wording, is the insistence that internal variables be explicitly separated from external ones and governed by formalized life-inspired dynamics, rather than compressed into a scalar bonus ^[4][3]. That is a testable difference. It is not tested in anything a reader can see: the supplied text stops at the reference list, with no experiment, no benchmark and no equation in it ^[19].

For anyone running hardware, the interesting citation is the aerospace one, on integrated system health management for mission-essential and safety-critical applications ^[14]. Telemetry of that kind is already collected. The proposal is about where it lives. If battery margin and thermal headroom stay in a supervisory monitor that vetoes the policy, the policy never learns to trade against them; if they enter the state as factorized variables with set-points, every such trade-off becomes part of what the agent optimizes, including the trades nobody specified. The reference list also reaches for a hierarchy of avoidance behaviours in a single-celled organism ^[15], which is the cheapest demonstration on offer that internal regulation yields graded behaviour without much of a brain. The paper adds neuromodulatory mechanisms as the second import, for context-dependent behaviour ^[6].

One thing worth pricing. Internal states are described as universally available ^[5]; the argument for them is not. The article costs $39.95, which is $6.96 more than a 30-day Nature+ pass at $32.99 that covers Nature and 54 other Nature Portfolio journals, roughly 21% more for one paper than for a month of 55 titles ^[8][17]. Both copies of the material supplied to us are the same abstract, same publisher, same URL ^[16], so none of it is corroborated by a second reader ^[18].