Towards a Science of Scaling Agent Systems (2512.08296v1)

Abstract: Agents, LLM (LM)-based systems that are capable of reasoning, planning, and acting are becoming the dominant paradigm for real-world AI applications. Despite this widespread adoption, the principles that determine their performance remain underexplored, leaving practitioners to rely on heuristics rather than principled design choices. We address this gap by deriving quantitative scaling principles for agent systems. We evaluate this across four diverse benchmarks: Finance-Agent, BrowseComp-Plus, PlanCraft, and Workbench. Using five canonical architectures (Single, Independent, Centralized, Decentralized, Hybrid) instantiated across three LLM families, we perform a controlled evaluation spanning 180 configurations with standardized tools and token budgets. We derive a predictive model using empirical coordination metrics, including efficiency, overhead, error amplification, and redundancy, that achieves cross-validated R^2=0.513. We identify three dominant effects: (1) a tool-coordination trade-off: under fixed computational budgets, tool-heavy tasks suffer disproportionately from multi-agent overhead. (2) a capability saturation: coordination yields diminishing or negative returns (beta=-0.408, p<0.001) once single-agent baselines exceed ~45%. (3) topology-dependent error amplification: independent agents amplify errors 17.2x through unchecked propagation, while centralized coordination contains this to 4.4x. Centralized coordination improves performance by 80.9% on parallelizable tasks like financial reasoning, while decentralized coordination excels on dynamic web navigation (+9.2% vs. +0.2%). Yet for sequential reasoning tasks, all multi-agent variants degraded performance by 39-70%. The framework predicts the optimal coordination strategy for 87% of held-out configurations, providing a predictive principle of agentic scaling based on measurable task properties.

• The paper presents a predictive framework that quantifies trade-offs between agent count, coordination overhead, and task complexity in LLM-based systems.
• It demonstrates that multi-agent system performance varies with task benchmarks, achieving gains in parallelizable domains while incurring costs in sequential tasks.
• The study highlights vendor-specific sweet spots and exposes limits due to error amplification and coordination inefficiencies, guiding optimal system design.

Quantitative Scaling Principles for Agentic Systems

Introduction

"Towards a Science of Scaling Agent Systems" (2512.08296) systematically investigates the interplay between agent count, coordination topology, model capability, and task characteristics in LLM-based agentic systems. Through 180 controlled experiments spanning four agentic benchmarks (Finance Agent, BrowseComp-Plus, PlanCraft, Workbench) and three LLM families (OpenAI, Google, Anthropic), the work derives actionable scaling laws for multi-agent coordination. The analysis identifies strong, domain-contingent effects that overturn the frequent assumption that simply scaling up agent count universally amplifies reasoning power. The authors introduce a predictive framework, parameterized by empirical metrics of coordination efficiency, error amplification, and protocol overhead, that anticipates optimal architectural choices and exposes fundamental resource ceilings for agentic collectives.

Agent and System Architecture

The paper delineates five canonical system architectures: Single-Agent System (SAS) and four Multi-Agent System (MAS) topologies—Independent, Centralized, Decentralized, and Hybrid. Each system is formally defined by task-agnostic mechanisms for agent reasoning, memory, action space, and inter-agent communication.

Single-agent architectures maximize context integration but lack parallel specialization and error redundancy; MAS variants decouple reasoning but incur message-passing overhead and lossy coordination. The authors introduce a comprehensive complexity taxonomy over LLM calls, memory, and communication, enabling strictly causal attribution of performance gains to architectural innovations rather than confounded implementation details.

Task Benchmarking and Coordination Topology

Four benchmarks represent complementary slices of the agentic workload spectrum: Finance Agent (parallelizable quantitative reasoning), BrowseComp-Plus (dynamic web navigation), PlanCraft (constraint-intensive sequential planning), and Workbench (structured tool execution). Selection is explicitly organized to stress-test coordination architectures across regimes of decomposability, sequential interdependence, tool density, and partial observability.

Crucially, the analysis distinguishes agentic from non-agentic tasks—validating that coordination only persists as a meaningful construction for environments requiring multi-step, adaptive reasoning under feedback and partial information. This task/architecture alignment precludes the misinterpretation of static benchmark ensemble gains as indicative of agentic scaling in dynamic domains.

Empirical Scaling Laws

A principal contribution is the derivation of a mixed-effects scaling model that outperforms naive architecture- or capability-only baselines. The model encodes interactions among model intelligence, task tool count, agent team size, computational overhead, coordination efficiency, redundancy, and observed error amplification. Its fit ($R^2_{CV}=0.513$) enables the prediction of MAS efficacy for unseen domains and architectures, quantifying the constraints and returns of agent scaling.

Three dominant effects are characterized:

• Tool-Coordination Trade-off: Multi-agent overhead interacts negatively with increasing tool density (interaction coefficient $\beta=-0.330$, $p<0.001$), often rendering MAS approaches counterproductive in tool-rich domains.
• Capability Saturation: Once SAS performance exceeds $\sim$45%, MAS infusion yields diminishing or negative returns due to coordination costs outweighing marginal diversity/resilience benefits ($\beta=-0.408$, $p<0.001$).
• Error Amplification Topology: Independent MAS propagates errors up to $17.2\times$ versus SAS, while Centralized orchestration constrains this to $4.4\times$ through validation bottlenecking. Architectural differences in verification and information bottlenecking are quantitatively mapped to generalized error trajectories across domains.

Task- and Vendor-Dependent Scaling

Performance improvements from multi-agent coordination are highly architecture- and domain-contingent. Centralized and hybrid MAS architectures yield the most substantial gains in structured, parallelizable domains (e.g., Finance Agent: $+80.9\%$) but often strongly degrade in sequentially constrained tasks (e.g., PlanCraft: $-70.0\%$). Decentralized configurations excel on dynamic web tasks requiring parallel exploration under high environmental entropy, but hybrid and over-coordinated systems succumb to compounded communication overhead and protocol complexity.

Figure 1: Average agentic performance increases monotonically with model Intelligence Index, but the magnitude and direction of MAS benefit are coordination- and domain-dependent.

Empirical results further reveal distinct cross-vendor architecture alignments: OpenAI models benefit most from centralized/hybrid protocols, Google models plateau under lightweight coordination with efficiency benefits, and Anthropic models are prone to coordination-induced variance and underperformance in high-overhead regimes. These differences point to architecture- and vendor-specific sweet spots, presumably rooted in differences in model communication bandwidth, reasoning-alignment, and context handling.

Figure 2: Cost-performance trade-offs show that maximal agentic performance is not universally aligned with economic efficiency, and optimal coordination patterns diverge systematically across model vendors.

Quantitative Coordination Analysis

Coordination efficiency and error dynamics are dissected with fine granularity:

• Reasoning Turn Scaling: MAS coordination induces super-linear scaling of reasoning turns with agent count ($T \propto n^{1.7}$), rapidly exhausting computational budgets and resource allocations. Beyond 3–4 agents, performance saturates or declines as per-agent reasoning bandwidth is lost to protocol overhead.
• Message Density and Success: Success rate plateaus logarithmically with message density; overly dense communication yields negligible marginal improvements and signals a phase transition to over-coordination.
• Redundancy vs Diversity: Redundant reasoning shows limited positive effect beyond certain synergy-bandwidth thresholds, supporting an operational trade-off between diversity-driven error reduction and diminishing returns from parallel duplication.
• Error Typology: The taxonomy distinguishes architecture-driven failures, e.g., logical contradiction, numerical drift, context omission, and coordination misalignment. Centralized architectures uniquely suppress context omissions and contradictions through bottleneck mediation.

  Figure 3: Relative MAS versus SAS performance across benchmarks exhibits high-variance, task-dependent scaling; multi-agent coordination collapses in domains requiring high sequential interdependence.

  

  Figure 4: Increasing the number of agents exposes sharp coordination limits: performance gains from multi-agent scaling vanish when coordination overhead dominates per-agent reasoning power.

Agent Heterogeneity and Model Mixing

The framework allows the characterization of heterogeneous collectives composed of LLMs of varying capability levels. For certain model families and topologies (notably, Anthropic in centralized orchestration), heterogeneity—deploying high-capability subagents coordinating via low-capability orchestrators—yields performance gains over homogeneous high-end ensembles, reflecting nontrivial interactions between role assignment, communication protocol, and base intelligence.

Figure 5: For Anthropic models, centralized teams with heterogeneous composition (low-cap orchestrator, high-cap workers) outperform homogeneous high-cap teams, an effect not observed for OpenAI/Gemini.

Benchmark- and Domain-Specific Coordination Efficacy

Scaling laws, error dynamics, and coordination efficiency are benchmark-specific. In structured domains (Finance Agent), MAS achieves strong performance through effective parallelization and redundancy; in high-sequentiality environments (PlanCraft), coordination cost quickly outweighs any ensemble benefit, and SAS remains optimal. These patterns are robust across agent architectures, vendor/model, and alternative complexity metrics.

Figure 6: MAS scaling advantage is maximized in parallelizable, structured benchmarks, with rapidly diminishing or negative returns in open-ended or highly sequential environments. Cross-family variance is dominated by differences in communication alignment.

Implications and Future Directions

The findings have critical implications for the design and deployment of agentic AI:

• Practically, the results cautions against blanket deployment of multi-agent collectives on complex tasks: optimal architecture selection must be data- and metric-driven, informed by task decomposability, tool density, and the empirical scaling law.
• Economically, MAS architectures are frequently suboptimal under fixed cost/latency budgets; token and resource consumption grows faster than performance, requiring cost-aware agent system design.
• Theoretically, the work frames collective agent behavior in analogy with complex adaptive systems—highlighting the emergence of phase transitions, critical points, and resource ceilings.
• For future development, the extension of these principles to fundamentally heterogeneous teams (across architectures, training objectives, or modalities), and the integration of specialized coordination protocols for tool-rich or embodied environments, remain open challenges.

Conclusion

This paper demonstrates that agentic system scaling is governed by quantifiable, domain- and architecture-dependent trade-offs rather than simple monotonic gains from collective scale. The framework—rooted in measurable empirical metrics—enables the principled, predictive design of agentic architectures, facilitating robust deployment while minimizing reliance on ad hoc heuristics. This work establishes foundational principles for a genuine science of agent system scaling, with substantial implications for both theoretical research and practical multi-agent system engineering.