# Full Text: FarmWorks: Decentralized AI Agents for Personalized Solutions

> Extracted from `2024_FarmWorks.pdf`

---

## Page 1

FarmWorks: Decentralized AI Agents for Personalized Solutions
Vladimir Baulin (ORCID: 0000-0003-2086-4271)
Alex Vyatkin (ORCID: 0000-0003-1306-4620)
Avel GUÉNIN—CARLUT (ORCID: 0000-0001-8239-7264)
Daniel Friedman (ORCID: 0000-0001-6232-9096)
John Bolt (ORCID: 0009-0002-0878-5546)
Stefan Falkenstein (ORCID: 0009-0008-5054-4969)
Parishrut Jassal (ORCID: 0009-0008-4351-805X)
Celio Trois (ORCID: 0000-0002-7386-9749)
Jonathan Minchin (ORCID: 0000-0001-8152-1796)
v1 ~ September 12, 2024.
Additional information (budget and more) and all contacts: blanket@activeinference.institute
Problem Statement
Climate change intensifies agricultural challenges, requiring more and more advanced technological
solutions. Small farmers increasingly rely on technical assistance, which is becoming centralized, dominated
by large agricultural corporations and governments imposing sophisticated pre-designated solutions. As AI
proliferates within these centralized solutions, diseases mitigation methods, climate credits, government
subsidies, and regulations risk monopolizing farmers' activities. This tendency, amplified by AI development,
 threatens to undermine farmers' autonomy and limit their ability to make independent decisions, converting
them into consumers of centralized technological solutions.
Project Objectives
We propose to develop FarmWorks — a platform for human-AI interaction in agriculture that enables
personalized, farm-scale solutions while resisting power concentration associated with centralized AI
systems. FarmWorks addresses the above challenges by providing an open source decentralized AI-powered
agricultural platform that empowers individual farmers with cutting-edge technology while preserving their
autonomy and promoting sustainable practices. By integrating real-time data collection, edge computing, and
Active Inference models, FarmWorks enables farmers to make informed decisions tailored to their specific
contexts (for example, integrating humidity data and epidemiological models to assist farmers with
remediative and anticipatory treatments for mold).
The platform's community-driven approach fosters knowledge sharing and collective innovation, while its
ethical framework ensures that AI deployment aligns with human values and ecological sustainability.
FarmWorks thus offers a comprehensive solution that not only enhances agricultural productivity but also
promotes environmental stewardship and farmer empowerment in the face of global agricultural challenges.
Key components of Farmworks include:
1. Decentralized Data Collection and Processing
​
Customizable on-site sensor networks for real-time data gathering
​
Local data integration and AI model training to preserve privacy
2. AI-Assisted Decision Support
​
Virtual assistant for natural language interaction with farmers
​
Active Inference models for personalized, context-aware recommendations
3. Community-Driven Innovation Network
​
Social platform for farmers to share strategies and insights
​
Collective problem-solving and best practice identification
4. Ethical AI Deployment
​
Community-developed ethical framework

## Page 2

​
Transparent AI decision-making processes
​
Balancing AI assistance with human autonomy
5. Promoting Sustainable Agriculture
​
Optimizing decentralized farm management through data-driven insights
​
Supporting transition to regenerative practices
This approach addresses power concentration by centralized AI via:
​
Keeping data and decision-making power with individual farmers
​
Fostering a collaborative network and peer-to-peer technological support rather than relying on
centralized solutions from governments and corporations
​
Ensuring transparency and ethical considerations in AI deployment
​
Empowering farmers with personalized AI tools while preserving their agency and creativity
FarmWorks can serve as a model for decentralized AI applications in other fields, demonstrating how to
harness AI's benefits while mitigating risks of power concentration.
Implementation
FarmWorks establishes a platform for human-AI interaction in agriculture, where local data drive
personalized solutions tailored to each individual farm. This approach offers multiple strategic options and
counteracts the risk of power concentration inherent in centralized AI systems, ensuring that farmers not only
retain control and autonomy over their agricultural practices, but become co-creators of personalized
solutions that can be shared to other farmers. Further, this approach can be replicated across various fields,
such as decentralized science and strategies to mitigate the accumulation of power in broader contexts.
We present FarmWorks with its setup of sensors and the community of interconnected farmers with the
creation of agents for predicting diseases, crop harvest predictions, etc. It is important to realize that this
proposal can be replicated in other fields, showcasing the broader potential of this approach in terms of
digital trust of AI-empowered humans.
We aim to augment this decision power of farmers in the setting of decentralized AI and farm-scale
sensemaking and decision-making. We aim to support farmer practices through data-guided decisions and
consensual knowledge sharing (rather than reliance on centralized AI systems). A key objective is to align
farmer expectations and requirements with an advanced recommendation system based on Active Inference
principles. This system takes into consideration harvest predictions, meteorological factors, and scientific
disease models, creating an iterative feedback loop provided by Active Inference. The FarmWorks
development and implementation will be divided into six Work Packages (WP), described below, whose
iterations are shown in Figure 1.

## Page 3

WP1: Real-Time Data Collection Module
WP1 is focused on the integration and deployment of a fully customizable network of on-site sensors and
cameras that collect critical environmental and crop data and adapting to specific field conditions. This
network is based on developments of open-source projects Vin-Q and ROMI. The data collected includes soil
moisture, temperature, pH, conductivity, humidity. The system is designed to operate under any conditions,
continuously collecting and transmitting vital agricultural information to a local server autonomously,
independent of connectivity or electric power. The sensors are connected to the local data server via
communication technologies including 4G, 5G and LoRa networks working simultaneously. The flexibility
of these communication protocols allows the system to function effectively even in remote areas with limited
connectivity. LoRa, in particular, offers long-range, low-power communication, making it ideal for
continuous data transmission across large fields or challenging terrains. Furthermore, the sensors and
gateways are powered with solar panels to avoid dependence on power cuts. By focusing on decentralization
in in-field processing, WP1 reduces dependency on centralized solutions such as cellular networks and
closed-source AI models, providing monitoring system sthat continuously collects data in the fields.
Deliverables: 1. Design of affordable network of customized sensors that can communicate with a gateway
via 4G-5G connection and LoRa network. 2. A fully operational real-time data monitoring and alerting
system that integrates with IoT edge devices and securely transmits and processes data from diverse sources.
3. Installation in Vin-Q farmers pilot fields (20 fields in Catalonia region). Key persons: Jonathan Minchin,
Vladimir Baulin
WP2. Decentralized Data Integration and Processing Module
The FarmWorks’ platform foundation is built on the automatic real-time data collection, processing, and
analysis of data from field sources. Each farm stores locally collected data from several boards and it is
equipped to handle in-field analytics capable of running AI and machine learning models. Apart from
avoiding centralized power accumulation, such setup reduces the need for centralized data storage and
processing, by localizing data processing.
Decentralized data-processing modules will be created based on the Vin-Q platform, as it offers consistent
integration of various open source tools: Apache Superset, Zeppelin, Kafka, Spark, integrating data from
different sources such as traditional databases (e.g. PostgreSQL, MySQL, SQLite) as well as distributed (e.g.
Apache Druid, Hive). This data is curated and organized into a Distributed Database, which forms the
foundation for advanced analysis in-situ. It is also designed in such a way that a curated database that is
needed for decision support is composed from preregistered datasets.
The module operates on a decentralized model, where data collection and processing will be done in an
instance running locally at each farm, enhancing privacy and security. This decentralized approach allows AI
agents to read on local data and contribute updates to a global model without sharing raw data. The
decentralized architecture ensures that all training data remains on each farm, with only model updates that
can be optionally shared with other users to combine in a larger model. This method not only protects data
privacy but also improves the efficiency and accuracy of the models by leveraging local, highly personalized
context-specific information tailored by the intentions and the needs of the user.
The module includes a visual interface that enables users to interact with, explore, and visualize data through
custom dashboards. These dashboards serve as a centralized hub for accessing real-time and historical
information about farm conditions, soil health, and more. Users can quickly identify trends, anomalies, and
potential issues using visual tools like graphs, calendar, charts, and maps. The interface allows farmers to
make data-driven decisions about irrigation, fertilization, and other practices, ultimately enhancing
sustainability and productivity. This module will provide data availability for customized AI agents to train
and infer the decisions based on highly adaptable Active Inference models.
Deliverables: 1. Integration of real-time data from various sources, such as IoT sensors, weather updates,
and user inputs (month 1 - 10). 2. Make data available for consumption by bots and Active Inference models
(month 10 - 13). 3. A user-friendly AI-powered interface with customizable dashboards that provide

## Page 4

real-time and historical data visualization, supporting informed decision-making in regenerative agriculture
(month 13-24).  Key persons: Celio Trois, Vladimir Baulin.
WP3. Virtual Assistant for AI-human Interaction
The platform integrates virtual assistants designed to facilitate seamless interaction between farmers and AI,
making complex data and insights more accessible and provide a control of farmer of the agents seemingly.
The virtual assistant functions realized as a chat bot (e.g. Telegram or other platform) with an intuitive
interface that allows farmers to engage with the system in a natural, conversational manner. Through the
farmer information available in the personalized database that keeps all the records of the farmer fields ,
optionally using embeddings of LLMs, this assistant is trained on the specific data of each farm, including
soil analysis and sensor information, enabling it to provide essential feedback loop between the system and
give personalized recommendations and insights that will change with conditions.
The virtual assistant optionally leverages Open Source Large Language Models (LLMs) installed locally in
the platform with the function to interpret natural language queries, allowing farmers to easily interrogate the
extensive personalized Knowledge Database. It tailors responses based on the unique conditions of each
farm, offering evidence-based practices that have proven successful in similar contexts. Beyond answering
questions, the assistant can help farmers visualize current field conditions, predict future developments, and
understand the impacts of climate change and weather events on soil health through interacting with AI
agents processing the data of the farmer field.
Additionally, the virtual assistant supports the setup of monitoring systems, enabling farmers to establish
custom alerts based on specific conditions, such as weather projections or soil moisture levels. This proactive
approach allows for timely interventions, helping farmers optimize their practices and make informed
decisions while maintaining a sense of digital trust and full control over localized data, without inadvertently
sharing their decisions or data with government authorities, big tech firms, or other stakeholders.
Deliverables: 1. A virtual assistant that enables natural language interaction with the AI-driven platform
implemented in Telegram, providing personalized insights, recommendations, and monitoring capabilities to
support farmers in making informed, data-driven decisions. 2. Integration with decentralized platform and AI
agents through API. Key persons: Design: Jonathan Minchin, Vladimir Baulin, Technical implementation of
the front end of interaction: Stefan Falkenstein, Daniel Friedman  
WP4. Active Inference Models
The project seeks to develop generative state space models that leverage Active Inference to elevate the
interaction between IoT data, AI agents and farmers. These will include predictive and causal wine-grape
epidemiological models, models for recommending interventions to the farmer, and models of other farms.
Active Inference models leverage variational free energy and expected free energy minimization to conduct
approximate online Bayesian inference and planning. In the context of FarmWorks, this means that the AI
agents will be designed to perpetually learn from and predict data gathered through soil sensors, weather
stations, interactions with the farmer, and interactions with other farm’s models via the Social Data Network.
In particular, the interaction between farmers and their AI agents forms a feedback loop that is central to the
platform's operation. As farmers dynamically evaluate the recommendations provided by the AI and take
action, the outcomes of these actions are fed back into the system, further refining the AI's models. This
continuous loop of action, observation, and adaptation ensures that the platform becomes more accurate and
personalized over time, aligning closely with the specific needs and conditions of each farm. Deep learning
and reinforcement learning models may be effective, however can have high computational cost and low
interpretability, exacerbating the risks of centralized AI. In contrast, Active Inference generative models
implemented in RxInfer.jl can be computationally efficient, interpretable, and operated in a decentralized
sense/decision-making setting.  
To support the implementation of these Active Inference models, the platform will use and extend RxInfer.jl,
an open-source probabilistic programming interface for conducting variational inference. RxInfer is
specifically tailored for fast, online inference on edge devices, making it a critical enabling technology for

## Page 5

FarmWorks’s decentralized AI. Furthermore, the flexibility of probabilistic programming will allow for the
development and composition of a wide variety of proven models, from classic regression, causal modeling,
and time series analysis, to modern deep learning methods — all grounded in the data being collected from
sensors and farmer interactions.
Deliverables: 1. Development and deployment of the RxInfer.jl generative model library, which will be
adapted to the specific needs of agricultural decision-making. 2. Integration with data platform, 3. Integration
with Virtual assistant. 4. Interaction between various agents trained on different models or different
management styles (connection to social network based on data). Model Library: RxInfer.jl
Key persons: John Bolt, Daniel Friedman, Alex Vyatkin, Stefan Falkenstein
WP5. Social Network based on Data and Community-Driven Innovation
The platform extends its capabilities by fostering a new type of social network integrating data exchange in
existing networks of trust. Such social network enables farmers to share their personalized strategies and
insights within a community in form of well-defined protocols supported by extensive data, while also
keeping a sense of agency on their farming practices and the manner in which they share these practices.
This network facilitates a powerful exchange of knowledge, where individual successes can be collectively
analyzed, refined, and applied to solve common challenges. By participating in this community, users
contribute to a system of swarm intelligence, where collaborative problem-solving becomes a shared
endeavor, amplifying the effectiveness of individual actions through collective input.
Our main challenge in that regard is to overcome adoption issues related to the perception of the platform, as
potential users find it too removed or even antagonistic from their existing practices. To address this, we
propose to use a very simple QR code or NFC-based system, requiring embodied interaction, to transfer
permission for data access. Hence, the exchange of data access permission is likely to be perceived as an
extension of (rather than a threat to) existing social practices and network of trust, while the system handles
the technical data exchange in a transparent manner. Furthermore, this incentivizes farmers to seek
peer-to-peer interaction with new people, therefore extending prior networks of trust ; and ensures innovation
through the uneven diffusion of data. This system will however require a careful design process to ensure
ergonomy, accessibility, and adequacy to local social practices.
A core requirement for the adequacy of this system is to enable a coherent, but flexible coupling between
peers’ practices, hence enabling both the diffusion and selection of innovative techniques. In other words, the
users should be encouraged to keep an attitude of curious attention to the output of FarmWorks, as well as
the practices of other farmers, rather than either ignoring or blindly following them.  This design principle
will likely necessitate a careful design process based on extensive observation of and feedback from potential
users. However, we propose to articulate our basic design concept from existing research on social
organization and self-identity under Active Inference (see WP6). Overall, the process of community
validation will be designed to help propagate and test practices, hence building over time a collective
databank of techniques that are proven to work under similar conditions.
Deliverables: 1. The design of social interaction between farmers through AI agents, 2. Creation of a social
network platform automatized by AI agents that supports data and solution sharing, enabling swarm
intelligence, collaborative problem-solving, and the identification of best practices through
community-driven innovation. Key persons: Avel Guénin-Carlut, Jon Bolt
WP6. Social Integration of AI: Design, Ethics, & Sociology
The design of an agency-augmenting, autonomous AI system subtly but critically reframes the line of
reflection developed to describe and regulate the use of centralized AI systems. FarmWorks is indeed to
be understood as a resource augmenting the power of human cognition, rather than an authoritative system
replacing it. Concretely implementing this model while minimizing the potential for abuse raises an array of
design, ethical and sociological questions, which we aim to investigate during the project - both to inform
our development of FarmWorks, and to leverage this process into a more general insight in the social
integration of AI.

## Page 6

The first question we need to tackle is at the very foundation of our project: it is to motivate our decision to
take agency augmentation as the core expected quality of our AI system instead of more classical
consequentialist or deontological considerations. Although we believe that there exist fundamental reasons to
consider the respect of agency as a core tenet of ethics (notably to be found in virtue and enactive approaches
to ethics), our main argument here centers on the historical sociology of power. As outlined in the
introduction, the concentration of power usually goes with the domination of perspectives aligned with the
logic of power (Scott’s “view from above”). Following this line of reasoning, we propose to produce
academic work on the “longue durée” dynamics of the social integration of AI, addressing how
agency-augmenting systems may (or may not) help us prevent some of its potential harmful effects.
Furthermore, this project requires a demonstration of whether and how it is possible to implement AI system
in human social organization in an agency-respecting manner. To do so, we plan to rely on the Active
Inference literature on social organization and collective intelligence. A core intuition of this literature is that
human social behavior follows contextual prediction of one’s self-model, which is notably scaffolded by
social communication and the observation of peer behavior. A key parameter determining the level of human
agency in a given social context is the rigidity of the coupling between self- and peer-prediction. Hence, the
core design principle we propose to follow is to keep a coherent but flexible coupling between humans users’
self-model and FarmWorks, as well as between human users. We propose leverage this design principle to
design the interface of FarmWorks (see WP5), while using this process to articulate a more general
framework for agency-respecting AI.
Deliverables: 1. Comprehension research of the “longue durée” prospective history of social integration of
AI. 2. Comprehensive research studies to analyze the effectiveness of this decentralized ethical model,
providing insights and guidelines for fostering transparent and fair coexistence between humans and AI
agents in various societal contexts. 3. Implementation of tools and protocols that ensure clear visibility into
AI decision-making processes and data usage, including user-friendly dashboards and audit trails.
Key persons: Avel Guénin-Carlut, Parishrut Jassal
Proposed Methodology
Real-Time Data Integration (WP1+WP2)
The platform will integrate real-time data from various sources, such as sensors, weather updates, and user
inputs. This data will be transmitted to the platform, where it will be processed and analyzed using Active
Inference models developed by the Institute.
User Interfaces (WP3)
Users will have access to intuitive interfaces that allow them to:
​
Chat/informal interfaces and Data/Analytic/API options.  
​
Customize and tune AI models based on their specific needs and preferences.
​
Set personal preferences and ethical boundaries for AI assistance.
​
Visualize and interpret AI decision-making processes for transparency.
​
Provide feedback and corrections to improve AI performance and alignment.
Active Inference Models (WP4)
The Active Inference models will utilize variational free energy minimization to generate adaptive,
context-aware AI agents. These models will be designed to:
​
Using RxInfer.jl to make generative models which process real-time data from various sources.
​
Generate dynamic, personalized decision-making scenarios using discourse graphs.
​
Leverage relevant context to compute near-optimal policies through expected free
energy-minimizing algorithms.

## Page 7

Active Inference-driven Dynamic Decision Database (WP2 + WP4) with real-time sensor data (WP1)
The platform will generate a dynamic database of possible decisions using discourse graphs that will:
​
Contain personalized scenarios and solutions based on real-time data and user inputs.
​
Be continuously updated through iterative learning from real-time data.
​
Enable users to choose from a range of optimal solutions guided by the Active Inference models.
Community-Driven Innovation (WP5)
​
Users can share their personalized strategies within the community, allowing for swarm intelligence,
collaborative problem-solving, Valorization of effective strategies by the community.
​
Development of a social network based on data and solutions, fostering collective intelligence.
Social and Research outcomes (WP6)
​
Developing qualitative and social science research to contextualize and communicate the work being
done, and how it relates to broader efforts to ameliorate risks of centralized AI.
Implementation Plan and Timeline
Year 1: Foundation and Core Development
​
Months 1-3: Project setup, team assembly, and detailed planning.
​
Months 3-9: Development of the core Active Inference models and real-time data integration.
​
Months 9-12: Initial implementation of user interfaces and dynamic decision database.
Year 2: Refinement and Community Building
​
Months 13-18: Iterative improvement of core components based on testing and feedback.
​
Months 19-24: Development of community-driven innovation tools and engagement strategies.
​
Continuous: Growing the user and developer community through workshops, and online events.
Evaluation and Success Metrics (examples)
1. Technical Metrics: Performance and scalability of the Active Inference models, Quality and
Quantity of models, Accuracy and consistency of data integration and decision-making
processes.
2. Community Engagement: Number of active users and contributors to the platform,
Diversity of participants in the community-driven innovation network.
3. Adoption and Impact: Number of farms/users adopting the platform and the variety of
applications, Measurable improvements in productivity and sustainability.
4. Public Perception: Media coverage and public discourse around decentralized AI, Surveys
assessing user trust and satisfaction with the platform.
5. Policy Influence: Citations or references to the project in policy discussions and regulatory
frameworks, Adoption of project-inspired principles in AI governance initiatives.
Risk Assessment and Mitigation Strategies
1. Technical Challenges in Integrating Real-Time Data with Active Inference generative models:
​
Mitigation: Leverage our expertise in sensor technology and data integration, and collaborate
with leading researchers in these fields.
2. Privacy and Security Vulnerabilities:
​
Mitigation: Implement rigorous security audits, bug bounty programs, and monitoring
3. Resistance from Established AI Industry Players:
​
Mitigation: Engage in constructive dialogue, emphasize complementary rather than
competitive positioning, and demonstrate unique value propositions.

## Page 8

4. Slow Adoption Due to Technical Complexity:
​
Mitigation: Prioritize user experience design, provide comprehensive documentation and
tutorials, and offer support through community forums.
5. Ethical Dilemmas in AI Decision-Making:
​
Mitigation: Establish an ethics advisory board, implement transparent decision-making
processes, and engage in ongoing public consultations.


---
*Extraction method: pymupdf*