Log-Primary Agentic Systems

• Log-Primary Agentic Systems are defined by the use of append-only logs to formalize agent state, coordination, and preference aggregation.
• They support deterministic replay and compositional agent design, fostering safe and efficient multi-agent experimentation.
• This architecture unifies probabilistic, economic, and computational principles to ensure robust, auditable AI alignment.

A Log-Primary Agentic System is a class of agentic architecture in which all agent state, coordination, causality, and, often, preference aggregation is formalized in terms of append-only logs. These systems treat the log not as a retrospective data source but as the primary substrate—defining the agent’s memory, enabling deterministic replay and forking, supporting complex compositionality via log transformations, and facilitating robust alignment protocols. Log-primary systems unify probabilistic, economic, and program-architectural principles, leading to new possibilities for safe, auditable, and self-improving multi-agent AI.

1. Formal Models and Log-Score Grounding

Log-primary agentic systems possess an explicit mathematical grounding in log-based preferences, utilities, or histories. A paradigmatic example is the representation of agents as outcome distributions $P_i:\mathcal{O}\to[0,1]$ with epistemic utility provided by the log-score $U(P_i,y) = -\log P_i(y)$. The agent’s welfare is thus $W_i(o)=\log P_i(o)$, establishing a direct correspondence between beliefs, utilities, and the log architecture.

Agent composition employs weighted logarithmic pooling: $$P_{\rm comp}(y) \propto \prod_{i=1}^n P_i(y)^{w_i},$$ which is the unique minimizer of $\sum_i w_i\,\mathrm{KL}(Q\|P_i)$ over $Q$, and aligns with geometric mean opinion pooling. In logit form,

$$\log P_{\rm comp}(y) = \sum_{i=1}^n w_i \log P_i(y) - \log Z,$$

thus reducing composition to log-linear operations (Lee et al., 8 Sep 2025).

2. Sharp Structural Results: Possibility Frontiers

These systems exhibit sharp boundaries in compositional possibility. For linear pooling ($P_{\rm lin}(y)=\sum_i w_i\,P_i(y)$) it is impossible for all agents to strictly increase their expected log-score utility; at least some agents experience loss. Furthermore, in binary outcome spaces, weighted log-pooling enforces an unavoidable trade-off: precisely, for $P_i$ with beliefs $(x_i,1-x_i)$,

$U(P_i,y) = -\log P_i(y)$0

with $U(P_i,y) = -\log P_i(y)$1, so strict mutual welfare improvement is impossible. In contrast, for $U(P_i,y) = -\log P_i(y)$2, mutually beneficial log-pooling becomes constructively possible (Lee et al., 8 Sep 2025).

These frontiers imply that log-primary architectures are not merely a matter of convenience—they formalize which configurations admit unanimous improvement, compositionality, and robust subagent integration.

3. Recursion, Cloning-Invariance, and Tilt Analysis

Log-primary systems are recursively compositional, supporting factorization into arbitrarily many sub-agents with invariance under cloning. Formally, decomposing $U(P_i,y) = -\log P_i(y)$3 into $U(P_i,y) = -\log P_i(y)$4 with distributed weights leaves the overall log-pool unchanged if

$U(P_i,y) = -\log P_i(y)$5

The welfare gap $U(P_i,y) = -\log P_i(y)$6 is continuous in total variation and the set of parents admitting unanimous improvement is open in the simplex—i.e., these compositional properties are robust to perturbations (Lee et al., 8 Sep 2025).

Tilt-based analysis constrains the space of decompositions. Suppose $U(P_i,y) = -\log P_i(y)$7; compatibility requires $U(P_i,y) = -\log P_i(y)$8. Taylor expansion at $U(P_i,y) = -\log P_i(y)$9 yields

$W_i(o)=\log P_i(o)$0

Since $W_i(o)=\log P_i(o)$1, not all agents can achieve strictly positive derivative, ruling out trivial duplications as a mechanism for unanimous improvement (Lee et al., 8 Sep 2025).

4. Log-Centric System Architectures

Log-primary architectures go beyond probabilistic aggregation and pervade agentic system design:

• Event-Sourced Agents: The append-only log $W_i(o)=\log P_i(o)$2 becomes the definitive state. Every action, fact creation, model call, or rule change appends a typed event—enabling deterministic reconstruction of the agent's working graph $W_i(o)=\log P_i(o)$3 via pure functional "folding" over the log (Nakajima, 21 May 2026).
• Deconstructed State Machines over Shared Logs: Systems such as LogAct decompose agents into state machines (Driver, Voters, Decider, Executor), each operating exclusively via typed log entries. Intention, voting, and execution phases are driven entirely from, and recorded in, the log, guaranteeing interruptibility, auditability, and deterministic replay. All side-effects on the environment are append-before-execute, ensuring the log's primacy for recovery, introspection, and semantic intervention (Balakrishnan et al., 9 Apr 2026).
• Forkable Logs for Isolation and Counterfactuals: Pragmas such as AgileLog introduce continuous and severed forks, allowing agents to branch execution with $W_i(o)=\log P_i(o)$4 resource costs, provide bidirectional isolation, and enable atomic promotion/merging of agentic write branches (Bhat et al., 16 Apr 2026).

System	Log Role	Isolation / Forking	Primary Motivation
Probabilistic Composition (Lee et al., 8 Sep 2025)	Preference and belief aggregation	Cloning, recursion	Agent alignment and unification
LogAct (Balakrishnan et al., 9 Apr 2026)	Execution history and coordination	Recovery, agentic veto	Reliability, audit, safety
AgileLog (Bhat et al., 16 Apr 2026)	Data stream substrate	Cheap, deep forking	Performance/logical isolation
ActiveGraph (Nakajima, 21 May 2026)	State and causality substrate	Branchable runs	Auditability, self-improvement

5. Lineage, Introspection, and Auditing

By making the log the single source of truth, log-primary agentic systems guarantee perfect causal provenance. In an event-sourced agent, each event captures explicit $W_i(o)=\log P_i(o)$5 pointers, so the entire derivation from initial goal $W_i(o)=\log P_i(o)$6 down to low-level model outputs is recoverable as a DAG of event causality. Artifacts such as claims, code, or decisions embed their lineage, granting full traceability with $W_i(o)=\log P_i(o)$7 pointer following in the stored log (Nakajima, 21 May 2026). Semantic recovery and introspection become possible by simply reading and querying the log for incomplete tasks, discrepancies, or failure points; for example, driver routines can prompt an LLM to generate optimized recovery actions using only the logged event history (Balakrishnan et al., 9 Apr 2026).

6. Operational Guarantees and Agentic Safety

Log-primary architectures confer strong correctness and recovery guarantees:

• Deterministic Replay: Given a log $W_i(o)=\log P_i(o)$8, replaying it always yields the same state and output, provided behaviors are deterministic functions of their inputs with no out-of-band side-effects (Nakajima, 21 May 2026).
• Fork and Replay Independence: Runs can be branched at any event $W_i(o)=\log P_i(o)$9, with replay of the prefix incurring no API or external calls, only in-memory folding. This enables honest counterfactual evaluation and efficient experimentation (Nakajima, 21 May 2026).
• At-most-once Execution and Recovery: Executor components guarantee that no intention is re-executed after a crash; semantic recovery Intention(s) are themselves mediated by the log (Balakrishnan et al., 9 Apr 2026).
• Safety and Veto: Actions must be approved by (possibly LLM-driven) voters over the log before execution, providing enforceable safety invariants and adversarial robustness with minimal impact on benign utility (3% drop vs. frontier models with dual-voter configurations) (Balakrishnan et al., 9 Apr 2026).

7. Multi-Agent Implications and Alignment Phenomena

Log-primary agentic systems crystallize implications for multi-agent architecture and alignment:

• Unification of Beliefs and Preferences: Log probabilities are treated as both beliefs and utilities, supporting a mathematically coherent blend of Bayesian, economic, and game-theoretic reasoning (Lee et al., 8 Sep 2025).
• Accounting for Antagonistic Subagents: Tilt analysis and explicit decomposition into “Luigi” (benevolent) and “Waluigi” (antagonistic) directions reveal that unrecognized adversarial modes cannot simply be suppressed. Manifesting and then suppressing Waluigi achieves strictly greater first-order reduction in misalignment than naïve positive-reinforcement strategies (Lee et al., 8 Sep 2025).
• Isolation for Safe Experimentation: Continuous and severed log forking allow safe parallel exploration, what-if analysis, as well as transactional integration of agentic outputs only upon explicit approval (Bhat et al., 16 Apr 2026).
• Auditability and Reversibility: Transformations in prompts, behaviors, and learned policies are all logged as first-class events, yielding transparent audit trails and simple rollback via log prefix replay (Nakajima, 21 May 2026).

---

Log-primary agentic systems represent a convergence of log-structured computation, probabilistic inference, and agentic compositionality. By making log-based beliefs, preferences, actions, and state transitions foundational, they provide the theoretical rigor, operational auditability, and practical modularity needed for advanced, robust, and auditable AI agents (Lee et al., 8 Sep 2025, Balakrishnan et al., 9 Apr 2026, Bhat et al., 16 Apr 2026, Nakajima, 21 May 2026).