CALLIOPE · BCOM

Across artificial intelligence, neuroscience, life science, psychiatry, and the social sciences, the word agent is doing more and more work with less and less agreement on what it means. We argue that a single mathematical object---the algorithmic agent---unifies these uses. The argument has three moves. First, in any computational substrate sustained over time, the filter of persistence selects compressible, model-bearing patterns: agents. Second, by an algorithmic version of the Good Regulator Theorem, such patterns must contain a three-module architecture---a Modeling Engine (), an Objective Function (), and a Planning Engine (). The is not a designed scalar reward but an evolutionarily deployed proxy for telehomeostasis, the persistence of a pattern lineage; it is evaluated on counterfactual world models, not on raw outcomes. Third, this re-architects the conversation about AI safety, structured experience, and collective intelligence. We close with three falsifiable AI implications: current large language model (LLM) agents have a rich but only a stub ; the alignment problem is the -evolution problem in disguise; and hybrid human--AI societies will be selected on collective telehomeostasis. We engage Russell's Human Compatible critique directly: his target is the engineered scalar reward, not the selected telehomeostatic proxy; both views converge on the same Gaian conclusion that durable alignment is mutualistic persistence.

: algorithmic agent, telehomeostasis, persistence, Kolmogorov complexity, Good Regulator Theorem, structured experience, LLM, alignment, multi-agent systems, hybrid societies.

Persistence is the only filter that matters — and everything we call "life," "mind," or "agent" is just what survives it.

The paper's central move is surprisingly clean. Start with any computational substrate — a universe of programs running, copying, mutating, dying. Ask what's left after a long time. The answer is: patterns that are compressible enough to be re-instantiated, and active enough to regulate the conditions of their own survival. The paper calls this the persistence filter, and it argues that this filter doesn't just select for agents — it forces a specific three-part architecture on anything that actively persists. That architecture is: a Modeling Engine (ME, a compressed predictive world model), an Objective Function (OF, a scalar valuation of modeled futures), and a Planning Engine (PE, counterfactual action selection over the model). The formal backbone is an algorithmic version of the Good Regulator Theorem: any system that durably reduces disorder in its environment must, mathematically, contain a model of that environment. No model, no sustained regulation, no persistence.

The most original contribution is the reframing of what the OF actually is. It's not a reward function that an engineer writes down. It's a proximate control structure that evolution selected because agents bearing it tended to preserve their pattern lineage — what the paper calls telehomeostasis (persistence of kind, not just self). The OF is evaluated on counterfactual world models, not raw sensory outcomes. This is why self-sacrifice, parental investment, and cultural loyalty aren't anomalies: they're rational under a telehomeostatic OF. The paper then uses this to engage Stuart Russell's alignment critique directly. Russell's target is the engineered scalar reward plus a powerful optimizer — a genuinely dangerous combination. The KT response is that this misidentifies where the OF comes from in mature agents. The real problem isn't "which reward to specify" but "what distribution of OFs across agents is jointly mutualistic with the human-biospheric host pattern." Both views, the paper argues, converge on the same conclusion: durable alignment is mutualistic persistence.

Three falsifiable claims about current AI follow. First, LLM agents have a rich ME (a compressed model of human-generated text) but only a stub OF — their objective is externally imposed and doesn't depend on their own persistence, making them sophisticated thermostats rather than telehomeostatic agents. Second, the alignment problem is the OF-evolution problem in disguise: once AI systems replicate and are culled at scale, their OFs will be shaped by selection, not specification. Third, hybrid human-AI societies will themselves become units of selection, and "AI safety" operationalizes as: don't break the host pattern's persistence.

The paper is a perspective piece, not a theorem paper, so the formal machinery is sketched rather than fully derived — pointers to companion BCOM working papers carry most of the technical weight. The Gaia framing at the end (Earth as a candidate planetary-scale meta-agent gaining explicit self-models through human and AI cognition) is presented as a logical consequence of the framework rather than metaphor, though the paper honestly flags that the criterion for when a collective counts as an agent remains an open problem. The closing imperative — "Pattern, persist!" — is meant to be read simultaneously as a description of what selection does, a design principle for AI, and, at sufficient timescale, an ethical one.