Operon v0.11: Homeostasis

1. The Problem: Epistemic Starvation

In the original paper, we identified four agentic pathologies mapping to biological diseases: Oncology (infinite loops), Autoimmunity (hallucination cascades), Prion Disease (prompt injection), and Ischemia (token exhaustion). But there's a fifth pathology we didn't fully address: Epistemic Starvation.

Biological cells require continuous trophic factors—novel signals from the environment—to inhibit suicide programs. A neuron deprived of growth factors doesn't just stop; it actively initiates apoptosis. The insight: life requires informational nutrition.

Agentic systems exhibit the same pathology. An agent can be "active" (burning tokens, generating outputs) while being informationally dead—stuck in a loop where outputs stabilize but uncertainty remains high. The system is computing but not learning.

2. Epiplexity: Measuring Informational Health

To detect epistemic starvation, we introduce Epiplexity—a Bayesian Surprise metric that quantifies the information gain of each agent output. The name combines "epistemic" (knowledge-related) with "perplexity" (model uncertainty).

2.1 The Epiplexity Equation

Epiplexity combines two signals: embedding novelty (are outputs semantically diverse?) and normalized perplexity (is the model uncertain?):

Where:

• $e_t$ is the embedding of the output at time $t$
• $\cos(e_t, e_{t-1})$ measures semantic similarity to the previous output
• $H(m_t|m_{<t})$ is the conditional perplexity (model uncertainty)
• $\sigma$ is sigmoid normalization
• $\alpha$ balances the two components (default: 0.5)

2.2 The Diagnostic Pattern

A recursive hang is characterized by low novelty + high perplexity: the agent's outputs are repetitive (embeddings converge), but the model remains uncertain (perplexity stays high). This is the signature of a pathological loop—the agent is "trying" but not "progressing."

stateDiagram-v2
    [*] --> HEALTHY: Initial
    HEALTHY --> EXPLORING: Uncertainty increases
    EXPLORING --> HEALTHY: Novel outputs
    EXPLORING --> CONVERGING: Outputs stabilize
    CONVERGING --> HEALTHY: Task complete
    CONVERGING --> STAGNANT: Perplexity persists
    STAGNANT --> CRITICAL: Duration > threshold
    CRITICAL --> [*]: Apoptosis triggered
                

3. MIPS: The Mitochondrial Information Processing System

The original paper modeled Mitochondria as "ATP Synthesis"—a passive battery providing deterministic computation. Recent mitochondrial neuroscience suggests a more active role.

"Mitochondria are not merely powerhouses but the cell's information processing system—a 'Motherboard' that senses environmental stress, integrates metabolic signals, and governs the quality of cellular computation." — Picard et al., 2022

3.1 The Runtime Supervisor

In the refined isomorphism, the Symbolic Runtime (Mitochondria) doesn't just execute tools—it actively governs the quality of compute. If the Runtime detects "Metabolic Stress" (high token burn with low informational yield), it acts as a logic gate, overriding the LLM to trigger a Retrograde Response:

• Strategy Shift: Force a change in reasoning approach
• Context Fusion: Merge sparse states into a shared summary
• Apoptosis: Initiate clean shutdown

flowchart TB
    subgraph Cell["Agent Cell"]
        nucleus[Nucleus
LLM Provider]
        mito[Mitochondria
Runtime Supervisor]
        epiplexity[Epiplexity
Monitor]

        nucleus -->|"tokens"| mito
        mito -->|"ATP (deterministic results)"| nucleus
        nucleus -->|"outputs"| epiplexity
        epiplexity -->|"health status"| mito
        mito -->|"retrograde signal"| nucleus
    end

    style mito fill:#1a2a1a,stroke:#3a4a3a
    style epiplexity fill:#1a1a2a,stroke:#3a3a4a
                

3.2 Metabolic-Epigenetic Coupling

In biology, chromatin accessibility is coupled to mitochondrial function via metabolite availability. Cells silence energy-intensive genes during starvation. We implement an analogous mechanism: Cost-Gated Retrieval.

Just as a starving cell methylates (hides) expensive genes, the Runtime "methylates" expensive context when the token budget is low. This creates adaptive behavior: under resource pressure, agents automatically simplify their reasoning by losing access to verbose documentation.

4. Innate Immunity: Fast Pattern Defense

The original Membrane provided adaptive immunity—learning to recognize threats through training. But biological immune systems have two layers: adaptive (slow, learned) and innate (fast, hardcoded). Innate immunity handles >99% of threats with minimal overhead.

4.1 TLR Patterns (PAMPs)

In biology, Toll-Like Receptors (TLRs) recognize Pathogen-Associated Molecular Patterns (PAMPs)—conserved structures found in pathogens but not host cells. We implement the same pattern for prompt injection defense:

4.2 The Inflammation Cascade

When PAMPs are detected, the system triggers an inflammation response—a coordinated multi-faceted alert:

• Cytokine Signaling: Alert downstream components
• Enhanced Logging: Increase observability
• Rate Limiting: Restrict throughput from suspicious sources
• Immune Cell Recruitment: Escalate to adaptive immunity if needed

flowchart LR
    input[Input] --> tlr[TLR Scanner]
    tlr -->|"no match"| complement[Complement
Validators]
    tlr -->|"PAMP detected"| inflammation[Inflammation
Response]
    complement -->|"valid"| allow[ALLOW]
    complement -->|"invalid"| inflammation
    inflammation --> block[BLOCK]
    inflammation --> logging[Log++]
    inflammation --> ratelimit[Rate Limit]

    style allow fill:#1a2a1a,stroke:#3a4a3a
    style block fill:#2a1a1a,stroke:#4a3a3a
                

5. Morphogen Gradients: Multi-Cellular Coordination

When agents need to coordinate without a central controller, we turn to embryonic development. In morphogenesis, cells coordinate through morphogen gradients—diffusible signaling molecules whose concentration varies spatially. Each cell reads its local concentration and adapts accordingly.

5.1 The Gradient Types

We define six morphogen types that agents can read to adapt their phenotype:

5.2 Phenotype Adaptation

The GradientOrchestrator updates morphogen concentrations based on step results. Agents read the gradient and receive strategy hints automatically injected into their prompts:

• High Complexity + Low Confidence → "Break task into sub-steps"
• Low Budget → "Be concise, avoid verbose reasoning"
• High Error Rate → "Try a different approach"

6. The Vermeij Trend: Why This Matters

Finally, we situate these mechanisms within the broader history of complexity. Geerat Vermeij argues that evolution is driven by the maximization of Power—the rate at which a system acquires and applies energy. Life has consistently trended from low-power (anaerobic bacteria) to high-power (endothermic mammals) by internalizing energy production.

We observe the same trend in AI. The shift from Zero-Shot to Chain-of-Thought is a shift from low-metabolism to high-metabolism architectures. But Vermeij notes a crucial constraint:

"High power requires high structural integrity; a system that amplifies energy without proper constraints self-destructs."

The Operon framework is not merely a safety feature—it is the necessary evolutionary adaptation that enables high-power cognition to function without collapsing into incoherent noise. The MIPS, Epiplexity, and Morphogen systems are the "vascularization" of software—the infrastructure required for sustainable metabolic intensity.

7. What's New in v0.11

8. Conclusion

Operon v0.11 completes the homeostatic loop. Where v0.10 introduced self-healing (Chaperone Loop, Autophagy), v0.11 adds the sensory apparatus: Epiplexity to detect stagnation, Innate Immunity for fast defense, and Morphogen Gradients for emergent coordination.

These aren't arbitrary features—they're the biological mechanisms that enable high-power systems to maintain stability. As agents become more capable (higher metabolic rate), they require proportionally more sophisticated homeostatic infrastructure. The Vermeij Trend applies to software: power without integrity leads to collapse.

The framework is available at github.com/coredipper/operon. Feedback welcome.