Skip to content

AgroAskAI: A Multi-Agentic AI Framework for Supporting Smallholder Farmers' Enquiries Globally

🕒 Published (v1): 2025-12-16 20:59 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AgroAskAI is a modular multi-agent framework for climate-adaptive agricultural decision support targeting smallholder farmers globally. It uses a chain-of-responsibility architecture where a central AgentManager dynamically selects and sequences specialized sub-agents, with an embedded Reviewer Agent providing iterative hallucination mitigation. On 20 qualitative queries, it outperforms both ChatGPT (GPT-5.1) and CROPWAT 8.0 across most evaluation dimensions.

Problem

Prior agricultural AI systems rely on single-agent models or use multi-agent frameworks only for static, pre-wired functions. They lack dynamic agent selection, internal governance to prevent hallucination, and real-time external data integration, making them unsuitable for context-specific, globally diverse climate adaptation queries from smallholder farmers.

Method

The system is built on Microsoft Semantic Kernel with GPT-4o as the underlying LLM across all agents. The pipeline follows a chain-of-responsibility pattern: 1. Prompt Agent — detects user language, collects input, forwards to ParseAgent, handles clarification requests. 2. Parse Agent — extracts intent, location, and time range into a JSON struct; iteratively requests missing fields via PromptAgent before proceeding. 3. AgentManager — dynamically selects which sub-agents to activate (WeatherForecast, WeatherHistory, Solution) using LLM-driven selection and termination strategies; aggregates outputs. 4. WeatherForecast / WeatherHistory Agents — call OpenCage (geocoding), OpenWeather (forecasts), and NASA POWER API (historical data) and return GPT-4o-summarized results. 5. Solution Agent — generates actionable recommendations using RAG over a curated external document repository plus weather data; iterates until Reviewer approves. 6. Reviewer Agent — evaluates outputs on clarity, technical soundness, fairness, and user intent alignment; fetches external sources to cross-check; can reject and re-trigger the AgentManager cycle.

Conversation history is logged end-to-end for traceability.

Key Contributions

  • Chain-of-responsibility multi-agent architecture with dynamic agent selection via a central AgentManager
  • Reviewer Agent as an internal governance layer capable of triggering full re-generation cycles
  • RAG integration with real-time APIs (NASA POWER, OpenWeather, OpenCage)
  • Parse Agent with iterative slot-filling for underspecified queries
  • Multilingual support (50+ languages) via LLM-native language detection
  • Full conversation logging per query for auditability

Results

Evaluation on 20 region-specific queries compared against ChatGPT (GPT-5.1, default settings) and CROPWAT 8.0:

  • Weather information: AgroAskAI listed forecasts in 16/20 runs and historical conditions in 16/20; ChatGPT achieved 7/20 and 3/20 respectively; CROPWAT provided rainfall metrics only with no temperature data
  • Soil conservation: AgroAskAI addressed in all 20 outputs with implementation instructions in 13; ChatGPT listed techniques in 16 with minimal detail; CROPWAT: 0
  • Livestock advice: AgroAskAI addressed in all 20 outputs; ChatGPT in 16; CROPWAT: 0
  • Farm adaptability: AgroAskAI provided small/large-scale step-by-step differences in 16/20; ChatGPT remained general; CROPWAT gave only timing metrics
  • Institutional support: ChatGPT cited programs (e.g., KCSAP, NDMA) in more outputs than AgroAskAI — identified as a gap
  • Reviewer Agent ablation: removal caused off-topic hallucinations (e.g., water conservation advice in response to a drought-crop-survival query)
  • Agent Manager ablation: confirmed correct selective activation (weather agents invoked only when query requires it)

Limitations

  • Output quality is bounded by the scope and accuracy of integrated external data sources (NASA POWER, OpenWeather, OpenCage)
  • Multilingual capability depends entirely on GPT-4o and may degrade for under-resourced dialects without fine-tuning
  • No deployment in low-connectivity environments; proposed mitigation (community computing hubs) is speculative
  • Evaluation is qualitative (20 queries, human-judged dimensions) with no automated metrics or statistical significance testing
  • Institutional resource integration lags behind ChatGPT's retrieval of support programs

Relevance to Harnesses / Meta-Harnesses

AgroAskAI's AgentManager is a direct instantiation of a meta-harness: it interprets parsed intent, dynamically selects which specialized sub-agents to invoke, accumulates their outputs, enforces termination conditions, and routes Reviewer feedback into re-generation cycles — all without hard-coded control flow. The Reviewer Agent's ability to reject outputs and re-trigger the full agent selection loop implements the retry/feedback pattern central to robust harness design. The Prompt-Parse-Manage-Execute-Review pipeline is a reproducible template for domain-specific multi-agent harnesses where correctness validation must be embedded rather than post-hoc. For researchers tracking harness architectures, this paper offers a concrete, deployed example of coordinator-driven orchestration with iterative self-correction in a high-stakes domain.