S1-NexusAgent: Hierarchical Research Agent

• S1-NexusAgent is a self-evolving hierarchical agent framework that automates long-horizon scientific research through structured dual-loop planning.
• It employs a dual-loop 'Plan-and-CodeAct' architecture for decomposing complex tasks and integrating heterogeneous tools with intention-aware orchestration.
• Experimental evaluations show state-of-the-art success rates across benchmarks, highlighting its robust continual learning and efficient context management.

S1-NexusAgent is a self-evolving, hierarchical agent framework designed for robust, long-horizon scientific research automation across disciplines such as biology, chemistry, and materials science. Distinct from conventional LLM-driven agents, S1-NexusAgent integrates a dual-loop “Plan-and-CodeAct” paradigm, intention-aware dynamic tool orchestration, object-reference-based sparse context management, and closed-loop continual learning via trajectory distillation. Experimental evidence demonstrates its state-of-the-art task success rates on benchmarks requiring complex multi-tool workflows and long-context scientific planning (Team, 2 Feb 2026).

1. Motivation and Context

S1-NexusAgent addresses major limitations observed in LLM-based scientific agents, including the instability of long-horizon planning, brittle goal maintenance, inadequate context management, and the absence of systematic knowledge retention. Contemporary scientific tasks often span dozens or hundreds of interdependent steps, requiring orchestration of specialized computational tools and iterative goal refinement—a setting in which prior agents demonstrate “goal drift,” tool integration failures, and no effective re-use of successful execution traces. S1-NexusAgent explicitly targets these deficits by architecting an end-to-end system around stable hierarchical planning, heterogenous tool integration, and the continual distillation of new “Scientific Skills” from successful trajectories (Team, 2 Feb 2026).

2. Hierarchical Plan-and-CodeAct Architecture

The control core of S1-NexusAgent is a dual-loop hierarchy:

• Outer Loop (“Plan”): The agent’s planner decomposes high-level scientific goals into a sequence of ordered subtasks using LLM reasoning and a library of previously distilled skills.
• Inner Loop (“CodeAct”): For each subtask, the agent iteratively generates executable code fragments or API calls, invokes cross-disciplinary tools, and observes the outcomes before accepting success or refining local strategies.

The dual-loop decoupling ensures that global coherence is maintained even during extensive local trial-and-error. Local failures or tool errors within the inner execution loop do not trigger global replanning, but upon persistent failure or subgoal completion, the outer loop is invoked for further decomposition and re-alignment. This structure is formalized as follows:

Algorithm 1: Dual-Loop Control

Input: Task T, Tool library M, Skill library S P₀ ← OuterPlannerLLM.plan(T, S) for k = 0,1,... (statusₖ, interpₖ) ← InnerExecutor.run(Pₖ, M) if statusₖ == SUCCESS: return interpₖ else: Pₖ₊₁ ← OuterPlannerLLM.replan(T, Pₖ, interpₖ) return “Failed”

The architecture isolates subtask context within object references and compresses results, optimizing for both modularity and scalability in large contexts (Team, 2 Feb 2026).

3. Model Context Protocol and Heterogeneous Tool Integration

S1-NexusAgent enforces a Model Context Protocol (MCP), a strict JSON-schema which all integrated computational tools (simulation packages, APIs, ML models) must implement. MCP standardizes:

• Tool name, version, and semantic API signature.
• Typed input/output arguments, units, and valid ranges.
• Natural language description of its purpose.

By translating all tool interfaces into MCP’s canonical form, the agent is agnostic to implementation details and can flexibly reason about tool selection and chaining. This unification enables intention-aware dynamic retrieval: subtask-level “intention embeddings” are computed and compared (via cosine similarity) with a prebuilt embedding index of thousands of tools, allowing rapid, on-demand tool selection and hot-plug integration. Faiss-based indexes underpin large-scale similarity search, and a cache is maintained for the most recently used tools (Team, 2 Feb 2026).

4. Object-Reference-Based Sparse Context Management

To address the intrinsic limitations of LLM context window size and prevent semantic interference across long scientific trajectories, S1-NexusAgent employs object-reference-based sparse context management. Each subtask context is represented externally (e.g., “OBJ_1234” handles) and only compressed summaries (typically 50–100 tokens) of completed subtasks are preserved in the agent’s active working memory. This achieves compression ratios of 10–20×, with salient tool outputs (e.g., critical numerical data or visualization URIs) retained for global planning and future skill distillation. Such isolation and compression drastically reduce context pollution and maintain stability across lengthy research workflows (Team, 2 Feb 2026).

5. Critic Agent, Scientific Skill Distillation, and Self-Evolution

After each dual-loop execution, an independent Critic Agent evaluates the entire trajectory according to scalar criteria:

$$\text{Score}(\tau) = \alpha \cdot (\text{success\_rate}) + \beta \cdot (\text{novelty}) + \gamma \cdot (\text{resource\_efficiency})$$

where $(\alpha, \beta, \gamma)$ are empirically tuned on historical data. Trajectories scoring above a threshold $\theta$ are “skillized”: distilled into parameterized, reusable Scientific Skills that are wrapped via MCP and added to the skill library. This process implements closed-loop continual learning—the agent systematically accumulates high-yield research paths as callable routines, improving with each batch of tasks and enabling generalization to unseen scenarios (Team, 2 Feb 2026).

6. Experimental Evaluation and Comparative Performance

Evaluation was conducted on three authoritative, tool-intensive benchmarks:

Benchmark	S1-NexusAgent	Top Baseline	Delta
BioMini-Eval	42.4%	SciMaster 36.1%	+6.3
ChemBench	55.2%	ChemCrow 47.8%	+7.4
MatSciBench	48.9%	HoneyComb 43.0%	+5.9

Gains are statistically significant ($p < 0.01$, paired bootstrap). The dual-loop architecture outperforms ablated, non-hierarchical agents by 4–6 pp on average. The automatic skill distillation mechanism translates to a 10% relative improvement over repeated tasks, demonstrating robust cross-task and cross-domain generalization (Team, 2 Feb 2026).

7. Current Limitations and Prospective Directions

Documented limitations include:

• Dependence on the scope and precision of the initial tool library and MCP wrappers.
• Computational cost from on-the-fly tool embedding and runtime installation.
• Manual tuning of Critic Agent weights $(\alpha, \beta, \gamma)$.

Proposed extensions include RL-driven Critic learning, parallel multi-agent inner loop execution, multimodal (e.g., imaging, spectra) extension of MCP and CodeAct, and automated discovery/generation of new tool wrappers. These directions aim to increase skill learning efficiency, parallelism, and support for broader data modalities (Team, 2 Feb 2026).

---

S1-NexusAgent defines a notable advancement in AI systems for scientific research, establishing an empirically validated architecture for robust, self-improving, and context-efficient orchestration of large-scale, tool-centric, multidisciplinary workflows (Team, 2 Feb 2026).