AgriAgent: Contract-Driven Planning and Capability-Aware Tool Orchestration in Real-World Agriculture

TL;DR

AgriAgent addresses the challenge of performing diverse, real-world agricultural tasks under multimodal inputs and incomplete tool ecosystems by introducing a router-based two-level architecture: System-1 for fast-path, modality-grounded reasoning and System-2 for contract-driven planning and capability-aware tool orchestration. System-2 encodes task requirements as explicit, verifiable contracts and leverages a centralized Tool Hub with dual matching protocols (TDI/TOCI) plus a ToolMaker for on-demand tool generation, enabling robust long-horizon execution with verification and evidence aggregation. The work demonstrates that contract-driven planning improves structural correctness and reliability on complex tasks, while the fast-path System-1 with modality-aware processing yields strong performance on simpler inquiries, collectively delivering improved robustness and scalability in realistic agricultural scenarios. By providing explicit contracts, dynamic tool generation, and verifiable execution, AgriAgent advances reproducible research and practical deployment in open, evolving agricultural tool ecosystems.

Abstract

Intelligent agent systems in real-world agricultural scenarios must handle diverse tasks under multimodal inputs, ranging from lightweight information understanding to complex multi-step execution. However, most existing approaches rely on a unified execution paradigm, which struggles to accommodate large variations in task complexity and incomplete tool availability commonly observed in agricultural environments. To address this challenge, we propose AgriAgent, a two-level agent framework for real-world agriculture. AgriAgent adopts a hierarchical execution strategy based on task complexity: simple tasks are handled through direct reasoning by modality-specific agents, while complex tasks trigger a contract-driven planning mechanism that formulates tasks as capability requirements and performs capability-aware tool orchestration and dynamic tool generation, enabling multi-step and verifiable execution with failure recovery. Experimental results show that AgriAgent achieves higher execution success rates and robustness on complex tasks compared to existing tool-centric agent baselines that rely on unified execution paradigms. All code, data will be released at after our work be accepted to promote reproducible research.

Figures (4)

• Figure 1: A Router first analyzes the user query and routes it by task complexity. Simple queries are sent to System 1 for fast-path QA without planning or tool use. Complex tasks are routed to System 2 for contract-driven planning and tool execution.
• Figure 2: System 1 (Fast Path). Multimodal inputs are processed by AgriGPT, AgriGPT-VL, and AgriGPT-Omni, optionally augmented by RAG (dense, sparse, and graph paths). Independent reasoning results are fused by Synthesis and finalized through Aggregation to produce the QA output. This path avoids planning and tool orchestration, enabling fast and stable responses for simple agricultural queries.
• Figure 3: System 2 (Complex Task Path). The plan/spec is refined through Supervisor Debate (critique--defend--revise) to fix structural errors and missing dependencies. Each plan node is then assigned a need contract in the Negotiation stage. Based on these contracts, the Tool Hub retrieves and composes tools from the Tool Pool via TDI/TOCI, while the Tool Maker generates new tools when required. Execution is validated to produce a verifiable result.
• Figure 4: Multi-agent roundtable discussion following by critique–defend–revise. An initial plan skeleton is generated in Stage 0 (Plan). In Stage 1 (Critique), a supervisor identifies missing, redundant, or invalid nodes. Stage 2 (Defend) justifies necessary components and resolves disagreements. Stage 3 (Revise) applies explicit edits (insert, replace, wrap, remove) to produce an executable plan, which is then passed to downstream tool orchestration.