Source-linked AI summary
Agentic World Analysis (AWA) - an alternative way to explore systems and support decision making
Yongchao Zeng, Alexey Voinov, Calum Brown, Tatiana Filatova, Mark Rounsevell
TL;DR
Complex socio-environmental systems make future analysis difficult because uncertainties, nonlinear interactions, and emerging human processes are hard to represent. The paper introduces AWA and implements it as WEGA, which uses expert agents to iteratively generate and audit scenarios. In the Netherlands nitrogen-crisis case, WEGA produced alternative pathways whose credibility depends on internal coherence and evidence grounding rather than calibrated predictive reliability.
Problem
Complex environmental systems coupled with human processes create uncertainties that existing modelling and expert-based approaches cannot fully handle.
Method
The paper proposes Agentic World Analysis and implements it through WEGA, using AI agents as experts to construct context, analyse decisions, explore scenario branches, and record the process.
Results
WEGA generated and evaluated alternative nitrogen-crisis pathways for the Netherlands, including a main storyline and an alternative branch involving monitoring restoration failure.
Takeaways & Limitations
AWA offers an alternative for exploring and comparing possible futures while exposing mechanisms, preconditions, failure modes, and analytical caveats.
Takeaways & Limitations
Scenario coherence and evidence grounding do not establish calibrated predictive reliability for projected tipping points, timelines, or recovery speeds.
Abstract
from arXiv · showhide
To address increasingly pressing sustainability challenges, various approaches have been developed to foresee possible futures, identify failure modes, detect vulnerabilities, and test potential mitigations. However, environmental systems are highly complex. Especially when coupled with human processes, the scale of uncertainties becomes intractable. To address this challenge, we propose a new approach - Agentic World Analysis (AWA)- combining the strengths of simulation modelling and expert elicitation. The concept of AWA is defined by three properties: 1) AWA uses an agentic AI system to mimic an expert panel that studies the world; 2) AWA projects futures iteratively through analysing scenario trees and learning from this analysis to improve decisions; 3) AWA is auditable. Based on these requirements, we implemented the World Engine by Generative Agents (WEGA) as a possible application of the AWA approach and demonstrated its functionality with a real-world case study: the Nitrogen Crisis in the Netherlands. WEGA autonomously constructed the context, identified key stakeholders and uncertainties, created expert agents, and generated future scenarios. As a result, two pathways from 2026 to 2041 were proposed, sharing a common assumption that social acceptance of nitrogen mitigation policies is low, while differing in how successful the restoration is according to the implementation of nitrogen data monitoring. The pathways are evaluated in multiple dimensions to assess their logical coherence and quality. The evaluation also actively exposes strengths and weaknesses to provide ways for testing the validity of the policies proposed. We discussed scaling up scenario analyses to enable massive pathway exploration, the trade-offs of using AWA and other approaches, and common concerns regarding AI systems.
1. Introduction
The paper introduces Agentic World Analysis as a response to the difficulty of analysing complex, uncertain socio-environmental systems with existing models and expert-based approaches. It combines simulation modelling, expert-centred analysis, and agentic AI to generate adaptive, auditable future scenarios.
- Motivation: Socio-environmental systems involve nonlinear feedbacks among human activities, ecosystems, economic development, social power, and long-term sustainability.Institutional shifts, public reactions, conflicts, and unexpected incidents may emerge spontaneously and be difficult to formalise.
- Motivation: Existing models remain limited in representation, adaptability, knowledge integration, interoperability, reuse, and integration across systems and case studies.Differences in abstraction strategies, scales, assumptions, documentation, and model-building lifecycles further hinder reuse and learning.
- Related approaches: Participatory modelling shifts attention toward stakeholder engagement, learning, communication, and solutions that participants find acceptable and actionable.This approach opens modelling processes to experts and affected ordinary people rather than focusing only on model construction.
- Related approaches: Expert elicitation uses domain experts’ knowledge and experience to explore uncertain futures, including knowledge that is difficult to formalise and emerging events.Unlike simulations, expert elicitation may lack consistent protocols, definite causal structures, traceability, and iterative mechanisms.
- Contribution: Agentic World Analysis integrates simulation modelling, expert-centred approaches, and agentic AI for adaptive scenario generation, evidence retrieval, reasoning, and decision support.AWA autonomously constructs context, makes assumptions, produces claims, generates reasoning chains, and explores scenarios for research or managerial purposes.
- Contribution: The paper introduces AWA, explains its implementation principles, demonstrates a working system with a real-world case study, and discusses trade-offs with existing approaches.It also reflects on the use of AWA in socio-environmental research.
2. Methodology
AWA represents complex-world analysis through expert agents that iteratively explore scenario branches and expose their reasoning. WEGA implements this approach through structured, multi-phase workflows whose steps and outputs are recorded for auditability.
- 2.1 Concept of AWA: AWA creates AI expert agents as proxies for an expert panel studying a real-world system affected by human activities and natural processes.The system can determine panel composition and domain representation, while agents may use existing models or create simulations when needed.
- 2.1 Concept of AWA: AWA projects plausible futures iteratively by generating scenario trees, exploring branches, and refining temporary outcomes through successive questions and time windows.This makes future inference more fine-grained and controllable than requesting only a single direct answer.
- 2.1 Concept of AWA: AWA is auditable because it exposes how relevant results are derived rather than reporting only final outputs.This requirement addresses biases, incentives, priorities, and knowledge limitations that can affect both human experts and AWA.
- 2.2 WEGA implementation: WEGA uses four major phases: initialisation, main loop, branch scenario reloading, and evaluation.Its features include analysis configuration, contextualisation, agent generation, branch exploration, result analysis, and evaluation.
- 2.2 WEGA implementation: During initialisation, WEGA gathers user inputs, builds contextual information, spawns stakeholder and expert agents, and produces an initial world statement.The initial world statement provides the background information shared by stakeholder and expert agents.
- 2.2 WEGA implementation: In each main-loop iteration, stakeholders reach a structured decision, expert agents analyse its impacts, and a synthesis agent generates alternative scenarios for branch selection.Expert outputs record impact assessments, causal mechanisms, and uncertainties; a scenario may be selected autonomously or by a human user.
- 2.2 WEGA implementation: WEGA records all steps and outputs from analysis setup through evaluation in a structured manner to support human examination and AI-assisted inspection.This operationalises AWA’s emphasis on auditability.
3. Case study
WEGA demonstrates AWA through a Netherlands nitrogen-crisis case study spanning 2026–2041, generating and evaluating alternative policy pathways. The selected pathways expose how social acceptance, monitoring integrity, and remediation assumptions shape future developments, while remaining scenario inferences rather than predictions.
- Case setup: The case study examines the Netherlands nitrogen crisis from 2026 to 2041 across three five-year deliberative iterations, using contextualization and stakeholder identification.The case was selected because it combines severe environmental impacts with diverse stakeholders and entangled social, economic, and political challenges.
- Main branch: WEGA generated policies addressing nitrogen transition, zone-specific reduction targets, and transparency before projecting alternative futures under different social-acceptance assumptions.The selected 2026–2031 scenario assumed low acceptance and included non-compliance, monitoring sabotage, falsified reporting, and intensified farmer–elite conflict.
- Main branch: In 2031, the selected Successful Restoration branch linked restored and independently audited monitoring to financial support, reduced sabotage, and more technical policy debate.The Structured Nitrogen Transition Accord provided a conditional framework, including a 30-day restoration condition and a 90-day pause on herd-reduction orders.
- Main branch: By 2036, the main storyline reached a stabilised but fragile equilibrium while testing whether legacy groundwater remediation could be accelerated through engineered intervention.The selected Engineered Acceleration scenario assumed remediation technologies could intercept and process subsurface nitrogen loads.
- Alternative pathways: WEGA also explored an alternative Collapse of Data Integrity branch, demonstrating that branch selection can produce materially different pathway histories.The evaluation cautions that stabilization in the carried-forward runs should not be interpreted as convergence because unselected branches included governance collapse, EU direct rule, and irreversible biodiversity loss.
- Interpretation and limits: The outputs are scenario inferences designed to reveal mechanisms, preconditions, and failure modes, not probability estimates or calibrated predictions.Specific policy designs are illustrative devices, and narrative coherence does not establish predictive reliability for projected tipping points, timelines, or recovery speeds.
4. Discussion
The discussion presents AWA and WEGA as modular, auditable approaches for analysing complex socio-environmental systems, while emphasizing distributed development and pragmatic limits on interpretation and reproducibility.
- Composability and distributed development: WEGA’s agent-centric workflow supports independently developing, modifying, and replacing agents without re-engineering the rest of the system.
- AWA’s properties: The evaluation agent enables self-reflection by assessing analytical processes across multiple dimensions and exposing insights, caveats, and vulnerabilities.
- Validation and auditability: AWA preserves intermediate and final outputs as accessible narratives, allowing inspection of evidence, assumptions, logic chains, and conclusions.
- Composability and distributed development: Composability supports distributed validation by allowing researchers to contribute competing expert agents with different assumptions, workflows, and perspectives.
- Composability and distributed development: AWA’s modular agents and interpretable outputs may lower barriers to collaboration and knowledge exchange across modelling communities.
Conclusions and open questions
AWA offers an alternative approach for exploring and comparing possible futures in environmental decision-making, while leaving open questions about its relationship with human experts and stakeholders.
- AWA uses AI agents to explore and compare possible futures for environmental management and decision-making.The authors present it as an alternative approach whose results may be influenced by agent skills or context.
- A central open question is how human experts and stakeholders should participate in AWA-supported decisions.The paper asks who should set goals, define study purposes, choose scenarios and decision trees, and determine whether proposed decisions benefit people.
Appendix A
Appendix A describes an auditable three-phase research workflow: Search, Read, and Write, which structures agentic analysis from information retrieval through report generation.
- The workflow is divided into Search, Read, and Write phases to streamline agentic analysis.Figure A1 presents this research process as embedded in the contextualization and expert agents.
- Search: The Search phase generates queries from multiple angles, ranks sources by semantic similarity, and filters contextually mismatched candidates.The final search output is a Markdown file containing metadata about the most relevant candidates.
- Read: The Read phase fetches candidate full texts and extracts useful evidence and claims into a JSON ledger.It prioritizes local files, Semantic Scholar links, unpaywalled DOIs, and finally abstracts and titles when fuller texts are unavailable.
- Write: The Write phase produces an academic-style report in which the agentic reasoning process is auditable.The auditable record covers identified gaps, causal logic, evidence, and the logic used to form conclusions, with review through a web-based interface.
Appendix B
Appendix B lists stakeholder decision sets generated by WEGA, including financial support tied to environmental milestones and temporary adjustments to herd-reduction administration.
- The proposed fund replaces the Supply-Chain EPR Levy with conditional grants for technology upgrades and debt restructuring.Grant conditions include verified environmental milestones.
- New administrative herd-reduction orders and zone-specific moratoria are paused during a 90-day negotiation window.Existing legal reduction trajectories remain in place while compliance tracking shifts to quarterly outcome-based pathways.
SNTA-2036-DEC-01
SNTA-2036-DEC-01 makes mandatory reductions and related triggers conditional on verified deposition progress, shifting compliance toward environmental outcomes.
- Mandatory herd reductions, state buyouts, and culling triggers are suspended if verified biannual deposition trajectories toward Natura 2000 critical loads are maintained.A proportional, co-managed adjustment protocol activates only when habitat-level deposition plateaus across two consecutive MRV cycles.
- Compliance metrics shift from livestock headcounts to emission-outcome metrics.
SNTA-2036-DEC-02
The Resilience Fund is accelerated and synchronized with CAP eco-schemes to finance verified agricultural technologies, while ringfenced capital addresses legacy groundwater contamination.
- Up to 70% co-financing supports precision feeding, closed-loop manure processing, and ammonia capture.
- 30% of the capital is ringfenced for legacy groundwater remediation, including riparian buffers, controlled drainage, constructed wetlands, and soil carbon restoration.
- Disbursements are milestone-tied but protected from retroactive clawbacks for non-systemic reporting deviations.
SNTA-2036-DEC-03
The proposal mandates recalibration and open-source publication of the JNAMR model while separating legacy groundwater nitrogen from current accounting pending hydrological validation.
- JNAMR model recalibration is mandated, with full source code, atmospheric dispersion assumptions, and weighting parameters published for open peer review.
- Legacy groundwater nitrogen is administratively decoupled from current operational accounting pending peer-reviewed hydrological validation.
- The RIVM, JRC, and EEA are assigned responsibility for publishing the model architecture and assumptions.
SNTA-2036-DEC-04
The Structured Nitrogen Transition Accord combines ecological monitoring, protected agricultural interests, conditional implementation, model review, financing, and temporary pauses in new reduction orders.
- SNTA-2036-DEC-04: Biannual ecological MRV checkpoints will measure habitat-level critical-load attainment at Natura 2000 sites.NGO and scientist coalition data on satellite NH3 tracking and biodiversity metrics becomes part of statutory EEA-MRV oversight.
- SNTA-2036-DEC-04: Independent Rural Impact Assessments and a two-thirds parliamentary majority are required for ecological set-asides or zoning overlays affecting active agricultural land.SNTA-compliant farms are shielded from automatic permit nullification, and livestock farming receives statutory cultural-continuity protections.
- SNTA-2036-DEC-04: The Structured Nitrogen Transition Accord replaces punitive paralysis with a conditional framework for legally compliant implementation.
- SNTA-2036-DEC-04: A 30-day restoration period requiring cessation of interference and reporting falsification precedes a 90-day conditional transition window.Compliance is verified through an independent third-party audit.
- SNTA-2036-DEC-04: JNAMR will stress-test deposition models and refine source attribution under guardrails against delay-driven baseline recalibration or directive overrides.
- SNTA-2036-DEC-04: The Supply-Chain EPR Levy is replaced by an agricultural transition resilience fund offering grants for technology upgrades and debt restructuring tied to verified environmental milestones.
- SNTA-2036-DEC-04: During the 90-day negotiation window, new herd-reduction orders and zone-specific measures pause while existing legal trajectories continue under quarterly outcome-based tracking.