Width, Memory, and Delay: A Resource Accounting for the Limits of Flat Multi-Agent Systems

Technical Summary: Width, Memory, and Delay: A Resource Accounting for the Limits of Flat Multi-Agent Systems

Problem Statement
The paper addresses a fundamental debate in the design of scalable multi-agent systems (ranging from robot swarms to large-language-model collectives): whether increasing the number of agents (population width) can indefinitely overcome performance limits, or if a qualitatively deeper, hierarchical organization is strictly required. This inquiry is motivated by a recent preprint [7] claiming that flat, homogeneous multi-agent systems possess an irreducible, population-independent "causal floor" on achievable error, which can only be lowered by hierarchical (nested-loop) organization. The authors argue that this claim conflates two distinct concepts: architectural flatness (lack of nested control loops) and memory flatness (lack of internal dynamical state). The paper seeks to replace this binary "flat vs. deep" debate with a quantitative resource model based on three distinct dimensions: population width ($N$), per-agent internal-model memory ($d$), and prediction across observation delay ($\tau$).

Methodology
The authors employ a controlled disturbance-rejection testbed grounded in classical stochastic control theory (minimum-variance control, the internal model principle, and Bode/waterbed constraints).

• System Model: A single scalar plant state $x[t]$ evolves under a common structured disturbance $d[t]$ (composed of narrowband tones and an AR process) and is observed by $N$ homogeneous agents. Agents receive delayed ($\tau$) and noisy ($v_i$) observations. Their control actions are averaged to form a collective input.
• Resource Definitions:
  • Width ($N$): The number of agents averaging independent measurement noise.
  • Memory ($d$): The dimension of the per-agent internal dynamical state (e.g., integrators, resonators) used to model the disturbance generator.
  • Delay ($\tau$): The latency between observation and control action, setting a prediction horizon.
• Controllers Tested:
  • Ladder: Flat homogeneous swarms with increasing per-agent memory (P, PI, and PI+Resonator banks).
  • Cascade: A designed two-loop hierarchical system (the "deep" baseline).
  • Kalman Frontier: An oracle-model, centralized minimum-variance controller serving as the computable optimum.
  • Adaptive: An oracle-free controller that learns disturbance frequencies online via periodogram peak-picking and phase-locked resonators.
• Evaluation: Performance is measured by steady-state Mean Squared Error (MSE). The study includes extensive experiments varying $N$, $d$, and $\tau$, alongside robustness checks against frequency mismatch, non-stationary disturbances (chirps and regime switching), mild nonlinearity, and spatially extended plants.

Key Contributions

1. Counterexample to the Strong Flat-Systems Theorem: The paper demonstrates that a flat, homogeneous swarm with matched per-agent internal models can achieve a lower error floor than a designed two-loop hierarchy when both possess equal per-agent memory. This refutes the claim that architectural hierarchy is necessary to lower the error floor; temporal depth can be achieved via dynamical (recurrent) memory rather than architectural nesting.
2. Quantitative Width $\times$ Memory Map: The authors establish that width and memory are not mutually interchangeable.
   • Width reduces the component of error due to independent measurement noise (scaling as $1/N$) but cannot reduce the structured residual caused by unmodeled disturbance frequencies.
   • Memory (specifically internal-model states) is required to cancel structured disturbances.
   • A strict equal-total-state-budget analysis ($N \cdot d = \text{const}$) reveals a threshold: below a certain signal-to-noise ratio, increasing width is beneficial; above it, investing states in memory (depth) decisively outperforms adding agents.
3. The Delay Theorem: The paper quantifies an irreducible error floor determined by the observation delay $\tau$ and the environment's unpredictability over that horizon. This floor corresponds to the variance of the multi-step-ahead prediction error of the disturbance plus the uncertainty from delayed, noisy sensing. This limit is verified against the optimal predictive controller (Proposition 2) to within 1%.
4. Cost of Adaptation: The study quantifies the performance gap between oracle-knowledge controllers and online adaptive controllers. While adaptive controllers approach the oracle baseline in stationary narrowband settings, they suffer from re-identification transients during fast regime switches, sometimes performing worse than robust memoryless loops.
5. Design Rules: Four practical guidelines are distilled for practitioners: (1) Count disturbance modes first (each needs internal-model states); (2) Add width only up to the point where noise reduction meets the structural floor; (3) Treat latency as a hard wall requiring prediction or latency reduction, not more agents; (4) Trade notch depth for robustness or adapt only if regimes outlive identification times.

Results

• Flat vs. Hierarchy: In Experiment E1-E3, a flat resonator swarm ($d=2$) achieved a lower MSE floor (0.058) than the best-tuned two-loop cascade (0.107) at equal per-agent memory, directly refuting the impossibility claim for flat systems.
• Non-Exchangeability: Experiment E4 showed that increasing $N$ cannot reduce error from a disturbance band not covered by the internal model (MSE remained $N$-independent for uncovered bands).
• Budget Analysis: In Block A, for a fixed total state budget, increasing memory depth ($d$) dominated increasing width ($N$) once a minimum SNR threshold was crossed. Below this threshold, measurement noise dominated, making width critical.
• Delay Limits: Block C confirmed that the MSE floor grows with $\tau$ and matches the semi-analytic Riccati floor of Proposition 2 to within 0.89%.
• Robustness: Sharp resonators provided deep suppression but were fragile to frequency mismatch. Adaptive controllers handled chirps well but failed under rapid regime switching due to identification transients.
• Generalization: The resource accounting held for spatially extended plants (12-node ring and 6x6 grid) and under partial observability, confirming the findings are not artifacts of the scalar testbed.

Significance and Claims
The paper claims to resolve the apparent paradox in multi-agent scalability by reframing the problem as "resource accounting" rather than architectural hierarchy. It argues that the "depth" required to overcome performance limits is not necessarily architectural (nested loops) but can be dynamical (internal models within agents).

• Scope: The authors explicitly state their claims are grounded in linear control theory with bounded nonlinearities and do not extend to chaotic or non-minimum-phase plants.
• LLM Relevance: While the testbed is a physical control system, the authors propose the resource triangle (width, memory, delay) as a hypothesis for cognitive collectives (e.g., LLM agents), suggesting that "memory" maps to context/retrieval and "delay" to reasoning latency. They posit that the empirical finding that a single strong agent can match a multi-agent workflow is explained by adequate internal modeling, while scaling helps primarily when it adds diversity (spectral coverage) rather than mere replication.
• Modesty: The authors avoid claiming to prove new impossibility theorems. Instead, they provide a counterexample to a strong existing claim and offer a sharper, classically-grounded resource model verified empirically. They acknowledge that their adaptive controller is a heuristic baseline and that distributed estimation (inter-agent communication) is a direct extension not yet fully explored.