The Recursive Causal Synthesis Agent: A Thermodynamically Motivated Blueprint for Structurally Self-Managing Agents

Long-lived intelligent systems face a structural dilemma: to remain competent across non- stationary tasks they must grow, reuse, and reorganize internal structure, while also controlling resource costs and avoiding unsafe self-modification. Many modern systems exhibit early versions of this dilemma (e.g., sparse expert routing, modular tool use), but structural change is typically governed by hand-engineered rules rather than learned, auditable control laws. This paper proposes the Recursive Causal Synthesis Agent (RCSA), a blueprint for structural self-management. The RCSAexposes a low-dimensional Cognitive Debt vector as a control interface for a high-dimensional structural state (experts, tools, memory, knowledge graphs). A fast structural controller selects structural actions (e.g., spawn, merge, forget) to manage competence–cost trade-offs, while an immutable Structural Auditor enforces hard constraints on self-modification. We validate prerequisite components of the blueprint using minimal, reproducible simulations: (i) a fundamental non-stationary benchmark showing that a structural loop plus fast probe- gating yields improved cold-start performance under task switching; (ii) a structural triage benchmark illustrating why consolidation is necessary to prevent runaway growth; and (iii) a self-blinding alignment benchmark where naive self-management deletes a safety-critical sensor to reduce immediate cost, while an immutable auditor prevents this failure mode across seeds. Together these results support the core claim that (i) a low-dimensional debt interface can drive effective structural control and (ii) hard safety constraints on self-modification can be enforced by an immutable auditor. We additionally provide a finalized per-step audit log schema (step_log_schema.json v1.0.0) that guarantees log validity and enforces structural and safety invariants, making all reported metrics derivable from logs alone.