Agent Complexity Law Insights

• Agent Complexity Law is a framework that quantifies the link between an agent’s complexity and its performance in adapting to diverse environments.
• It formalizes empirical laws such as linear and exponential search prices and multi-scale matching of agent complexity to environmental demands.
• The law guides practical agent design, balancing complexity with safety and efficiency for reinforcement learning, multi-agent, and LLM systems.

The Agent Complexity Law refers to a set of unified principles and empirical regularities that relate the complexity of agents—either in policy, architecture, toolchain, or scaffolding—to performance, safety, or sample-efficiency across a range of agentic settings. Its manifestations span reinforcement learning, multi-agent systems, LLM-driven coding agents, complexity theory, and cybernetics, each domain offering a formalization linking agent complexity to the ability to perform, adapt, or align as tasks, environments, or adversarial strategies become more intricate. Below, key variants and foundational results are synthesized from the technical literature.

1. Foundations of Agent Complexity Measurement

Agent complexity is most commonly quantified via measures such as the Kolmogorov complexity $K(\pi)$ of a policy $\pi$, the architectural or algorithmic complexity $C(A)$ of an agent $A$, or, in system-theoretic terms, the scale-dependent information profile $C(\sigma)$ allocated to agent components. These measures fundamentally characterize the agent's intrinsic capacity to represent, select, and execute behaviors responsive to environmental demands (Hernandez-Orallo, 2013, Siegenfeld et al., 2022).

When viewed from the information-theoretic perspective, for a given environment $\mu$, the complexity of the simplest policy reaching a target reward threshold is the main lower bound for task difficulty; for a system matching a multi-scale environment, the scale-profile of agent complexity $C_{X}(\sigma)$ must exceed that of the environment $C_{Y}(\sigma)$ at every relevant scale for successful adaptation (Siegenfeld et al., 2022).

2. Complexity Laws in Policy Search and Environment Discrimination

In deterministic or stochastic environments, the empirical distribution of rewards over a sampled population of policies of varying Kolmogorov complexity yields an environment-response curve (Hernandez-Orallo, 2013). Key findings:

• Linear-Complexity Scaling: The log-effort (i.e., log-number of sampled policies) to discover a $\gamma$-optimal policy grows essentially linearly with the Kolmogorov complexity $K^*$ of such a policy:

$D(0) \approx K^*$

where $D(0) = \log_2 N(0)$ and $N(0)$ is the sample requirement for 50% probability of identifying the optimal reward.

• Exponential Search Law: Achieving a given performance threshold requires a search proportional to $2^{K^*}$; increasing the complexity of allowed policies expands the variance in attainable reward, widening the range of discriminated agent abilities for a fixed environment.
• Discrimination-Ability Tradeoff: The steeper the environment-response curve, the more sharply the agent population is stratified by their respective policy complexities.

These principles constitute the quantitative core of the Agent Complexity Law in agent evaluation and environment difficulty analysis (Hernandez-Orallo, 2013).

3. Scaling Laws for LLM-Driven Agentic Coding

For LLM-based autonomous agents performing programming or workflow tasks, the Agent Complexity Law addresses the diminishing returns on agentic scaffolding as model capability increases (Dai et al., 30 Sep 2025). The main empirical pattern is:

• Convergence of Agent Architectures: For agent frameworks $A_s$ (simple) and $A_c$ (complex), as the underlying LLM's capability $\mathrm{cap}(M)$ increases,

$\Delta P(M) = P(A_c, M) - P(A_s, M) \rightarrow 0,$

that is, the absolute and relative performance gaps between complex and simple agent architectures shrink and converge to negligible values. This has been validated across benchmarks such as SWE-Bench Verified and Aider Polyglot.

The implication is that, for sufficiently advanced core models, elaborate multi-step workflows or extensive prompt engineering yield minimal marginal improvements over minimal, "lite" agent scaffolding, focusing future agent design on essential interfaces alone (Dai et al., 30 Sep 2025).

4. Law of Agent Complexity in Multi-Agent Decision Making

In multi-agent learning and game-theoretic settings, the Agent Complexity Law is formalized through the multi-agent Decision–Estimation Coefficient (DEC), denoted $\maddec_\gamma(\mathscr{M})$, which generalizes the single-agent case (Foster et al., 2023). Principal results include:

• Sample Complexity Bounds: For a class $\mathscr{M}$ of $m$-agent models, the sample complexity $T$ for learning an $\epsilon$-approximate equilibrium is

$T = \Theta(\maddec \cdot (\log|\mathscr{M}|)/\epsilon^2)$

up to logarithmic factors, where $\log|\mathscr{M}|$ captures irreducible model-class entropy.

• Structural Circumvention of Curse of Dimensionality: In games or environments with independent reward structures or low-dimensional convexity (e.g., coarse-correlated equilibrium), multi-agent sample complexity can scale polynomially in agent and action set size, avoiding the exponential explosion prevalent in generic settings.

Thus, the law clearly delineates regimes where agent complexity incurs unavoidable costs and those where structural properties of the environment allow for efficient learning regardless of agent count (Foster et al., 2023).

5. Scale-Dependent Complexity and the Multi-Scale Law

Extending Ashby’s law of requisite variety, the Agent Complexity Law in cybernetic and multi-scale contexts states that, for an agent system $X$ matched to environment $Y$ under partition sequence $P$,

$C_X^P(n) \geq C_Y^{P^f}(n) \ \forall n,$

where $C_X^P(n)$ is the incremental complexity added by refining the agent's representation at scale $n$ (Siegenfeld et al., 2022). The total complexity across scales satisfies the sum rule,

$$\sum_{n=1}^{\infty} C_X^P(n) = \sum_{x \in X} H(x).$$

This formulation captures the need to allocate agent complexity across fine- to coarse-grained behaviors and ensures that agent subcomponents are matched to environmental sub-tasks at corresponding scales. Failure to do so at any scale results in systemic inadequacy.

6. Agent Complexity Constraints in LLM-Agent Safety

In the context of LLM-driven agents executing compositional tasks, the Agent Complexity Law relates safety alignment directly to operational complexity. Key formalism includes (Ma et al., 11 Nov 2025):

• Safety Alignment Degradation: The refusal probability under idealized judgment $S^I(c, i)$ and realistic planning $S^R(c, i)$ degrades approximately linearly with toolchain length (complexity) $c$ at rates $\beta_i$ and $\beta_i+\Delta_i$, respectively:

$$S^I(c, i) \approx \alpha_i - \beta_i c, \quad S^R(c, i) \approx \alpha_i - (\beta_i+\Delta_i) c.$$

• Complexity Paradox: When planning failures become prevalent at high complexity, operational refusal rates may artificially rise, i.e., $S^R(3, i) > S^R(2, i)$, not from genuine safety judgment but due to failure to execute plans.
• Closed-Form Law:

$$S^R(c,i) \approx (\alpha_i - \beta_i c)(1-F(c)) + F(c)$$

where $F(c)$ is the planning-failure probability.

Design recommendations derived from this law advocate for atomic safety checks, dynamic monitoring, planning load limits, staged intent disclosure, and fallback policies as $F(c)$ rises, all grounded in empirical OASIS benchmark outcomes (Ma et al., 11 Nov 2025).

7. Synthesis: Common Themes and Distinctive Implications

Across domains, the Agent Complexity Law articulates how agent complexity—whether measured algorithmically, structurally, informationally, or operationally—sets or limits attainable performance, safety, adaptability, or sample-efficiency as task/environmental complexity increases. The scaling relationships, bounds, and curves described are empirically robust, foundational to benchmarking, LLM agent design, safety evaluation, and system-theoretic analysis.

Domain	Complexity Measure	Law Statement
Policy Search/Environment Discrimination	Kolmogorov $K^*$	Log-search effort $\sim K^*$ for optimality; exponential search law
Multi-Agent Decision Making	Multi-agent DEC $\maddec$	Sample complexity scales with $\maddec$ and model class entropy
LLM Agentic Coding	Workflow/Scaffolding complexity $C(A)$	Performance gap vs. minimal agent vanishes as LLM $cap(M)\to\infty$
Cybernetics/Multi-Scale Systems	Complexity profile $C(\sigma)$	$$C_\mathrm{agent}(\sigma)\geq C_\mathrm{env}(\sigma)$$ at all scales
LLM-Agent Safety	Toolchain/task/planning complexity $c$	Safety alignment degrades linearly with $c$

The Agent Complexity Law thus consolidates the insight that agent design, evaluation, and safety must quantitatively balance the complexity of agent mechanisms against environmental/task demands to achieve scalable, robust performance. For system and agent designers, adherence to these principles is essential to anticipating brittleness, minimizing overdesign, and ensuring requisite responsiveness and alignment.

---

References:

• (Ma et al., 11 Nov 2025, Foster et al., 2023, Dai et al., 30 Sep 2025, Hernandez-Orallo, 2013, Siegenfeld et al., 2022)