Multiple Agents for Hydrodynamics: Beyond the Single Context Window

When a single-agent system (SAS) receives a complex scientific task, something predictable happens: the context window fills with tool specifications and observation traces, and the available space for reasoning about each individual decision shrinks progressively. It's a known problem, and until now the usual answer has been to increase window size or compress history. A group of researchers proposes a different way out: decompose the workflow into specialized agents coordinated through a layered execution graph.

The paper, published on May 6, 2026 on arXiv under identifier arXiv:2605.01102, describes a prototype multi-agent system (MAS) applied to hydrodynamics. The domain is not accidental: hydrodynamic simulations combine heterogeneous data, numerous tool calls, and result synthesis that fit perfectly into the scenario where SAS begin to fail.

How the Layer Execution Graph Works

The core of the architecture is the Layer Execution Graph (LEG): an execution graph that the planning agent builds dynamically for each query. Rather than encoding control logic as rigid rules, the planner translates the natural language query into an execution topology guided by domain heuristics. This allows adapting the workflow to the nature of each question without rewriting the system logic.

Each specialist agent operates under a strict tool allowlist, reducing the risk of invoking resources outside its competence and facilitating auditing. The roles are clearly differentiated:

• Specialist agents: process concrete data classes in parallel within each layer.
• Consolidator agents: merge parallel outputs into compact summaries before passing to the next layer.
• Reporter agent: synthesizes the final response from the consolidated results.
• Provenance system: the runtime records each tool invocation, enabling full traceability of the reasoning.

The idea of consolidators between layers is especially interesting: instead of a single agent accumulating all historical context, each consolidator acts as a controlled-loss compressor, keeping what's relevant and discarding operational noise.

Why Role Separation Matters

The proposal is not new in abstract terms, multi-agent system literature has been exploring agent topologies for years, but its application to real scientific workflows has immediate practical value. In environments like fluid simulations, data combines time series, vector fields, numerical results, and configuration metadata. Mixing all of that in a single agent's context is not only inefficient: it introduces cross-domain interference that degrades reasoning quality.

Role assignment with complementary allowlists also addresses a production engineering concern: monolithic single-agent systems are difficult to debug when they fail, because the error could be anywhere in the workflow. With LEG, each layer is observable and reproducible separately.

Claude Sonnet 4.6 as the Base Model

All benchmarks, ablations, and stress tests in the study use Claude Sonnet 4.6 as the reference model, for both specialist agents and general-purpose ones. The authors don't explicitly justify this choice in the abstract, but the decision is coherent with the model's profile: good balance between reasoning capability and cost per token, which in an architecture with multiple agents running in parallel has direct impact on the system's economic viability.

The fact that the study uses an existing, documented model, not an experimental version, makes it easier for other teams to replicate the experiments or adapt the architecture to their own scientific domains.

Who This is Useful For

This work interests mainly two profiles. On one hand, engineering teams already building scientific pipelines on Claude and hitting the limits of the single-agent pattern: the LEG architecture offers a concrete reference model for restructuring those workflows. On the other, researchers in computational simulation, not just hydrodynamics but also climatology, structural engineering, or bioinformatics, who want to explore how LLMs can participate in analysis workflows without becoming a bottleneck.

The proposal doesn't solve all multi-agent system problems, coordination between layers introduces its own latency and configuration complexity, but the approach of dynamic graphs with intermediate consolidation points in a more robust direction than simply scaling the context window.

---

From our perspective, we appreciate that the study focuses on auditability and traceability, two aspects that agent prototypes often sacrifice for demonstrating capability. Having benchmarks reproducible with a publicly available model is a step toward more useful research for the community.