Education for the master and the agent cloud

The founding position of the Light Academy and XPower educational resources

Alexey Ananin and the Echelon group

26 September 2026

We are building our academies as an environment for the continuous development of intelligence, specialised agents and the people who work with them. Core intelligence connects fields of knowledge. An agent cloud helps track research, compare evidence, perform calculations and check solutions. The master sets the direction, organises the team and takes responsibility for what ultimately happens to people, equipment and their environment.

There is a practical reason for this position. Our work extends far beyond a single profession. It brings together biology, engineering, space, electricity and energy more broadly, architecture, economics, national security, human performance and the capabilities of human communities. Each field keeps changing. New findings emerge both within disciplines and where they meet.

An academy therefore needs to help people understand the whole picture, study more deeply where their own judgment is required and manage a continuously updated knowledge system. For us, professional work is undertaken by a combination of master, human team, intelligence, agents and measurement tools.

1 Why our academy has such broad boundaries

Consider lighting. A luminaire has electrical power, a spectrum, optics, thermal behaviour, controls and a service life. A room has geometry, surfaces, daylight and patterns of use. A person has a visual task, a daily routine, an age, individual characteristics and accumulated fatigue. An organisation has costs, operational requirements and responsibility for working conditions.

A single lighting project therefore requires several professional languages. An engineer designs the installation. An architect understands the space. A biology specialist frames questions about the effects of light on the body. An economist checks ownership costs. An operator assesses repairability and the ability to maintain the intended conditions over years. The quality of the combined solution depends on how well this knowledge is connected.

In our educational resources, human performance includes seeing, processing information, learning, acting accurately, recovering and sustaining effort. Individual productivity concerns the results of particular activities. By population productivity we mean a community's combined capacity to learn, create, maintain infrastructure and sustain quality of life. These concepts require clear measures and careful movement between scales. A laboratory finding in a small group still needs testing in a school, store, city or industry.

Energy and national security enter this picture through infrastructure resilience, access to resources, reliable communications and computation, and the preservation of expertise. Space adds constrained resources, autonomy, long service periods and costly failures. A principle from a space system may prove useful on Earth; its applicability must be checked under specific conditions.

The past offers experience of durability, operation and failure. The future requires scenarios. Frontier science lets us consider ideas still undergoing evaluation. The academy should identify the status of each contribution: reproducible finding, preliminary observation, model, hypothesis or an author's conjecture. Ambitious questions call for precision about evidence.

2 The knowledge stream exceeds one specialist's capacity

Our preliminary estimate is that roughly 350–600 scientific papers appear daily across the broad set of topics relevant to our work. This is the authors' working estimate of scale. Bibliometric confirmation would require a defined topic set, databases, counting period and rules for removing duplicates, including a preprint and its subsequent journal publication.

Even this preliminary range indicates the workload. Allowing ten minutes for a first look at each paper gives approximately 58–100 hours every day. This is an illustrative workload calculation. Examining methods, checking calculations and comparing earlier findings takes additional time.

Professional work also involves sites, people, measurements, design, agreements and the consequences of decisions. A person can understand their own craft deeply while physically missing important developments in adjacent fields. Occasional refresher training cannot fully resolve this problem.

Publication alone says little about practical value. A system must recognise repetition, experimental limitations, corrections and retractions, and compare conflicting findings. Knowledge needs continuous selection while preserving access to its original sources.

We expect intelligence and agents to undertake an increasing share of this work. Automated coverage of a broad stream requires checks on the completeness and quality of the monitoring itself. A list of processed documents does not establish understanding or reliable conclusions.

3 What we mean by training intelligence

Core intelligence in our model consists of one or more capable models that understand a task, connect disciplines and coordinate work. A specialised agent receives a function, access to relevant sources and tools, verification rules and limits of authority. Several agents can use the same underlying model while performing different tasks.

Training encompasses several processes. Developers train and update foundation models. We shape their professional working environment: validated materials, terminology, methods, calculation tools, examples, instructions and evaluations. Separate fine-tuning on prepared data is possible when justified and adequately resourced.

Adding a new paper to a knowledge base or supplying it in an agent's current context updates the information available to that agent. Changing the model's parameters requires a separate procedure. The academy needs to describe precisely what changed, which version is in use and which checks it passed.

A job title alone gives an agent insufficient preparation. It needs a task, sound inputs, suitable tools and a clear criterion for success. A lighting agent should work with measurements and calculations. An economics agent should expose assumptions and model sensitivity. A scientific-evidence agent should distinguish experimental results from the authors' interpretation.

An agent cloud is useful when responsibilities can be separated and their outputs reconciled. We relate the number of agents to task complexity, the cost of error and computational cost. A simple question may need one agent. A complex project requires several competencies and a separate check of their compatibility.

4 The academy must serve people and agents

Our academy has two audiences. People need a map of the subject: key concepts, causal relationships, important questions, common mistakes and an understanding of whom to involve. Agents need materials they can reliably find, interpret, apply and verify.

Each significant method should identify its sources, review date, scope, limitations and examples. Calculations should expose formulas, units, inputs and assumptions. Claims of an effect should explain measurement conditions and the comparison used.

People may learn through a short article, a site review, a conversation with a master, laboratory work or a real task. For agents, the same material may take the form of a structured record, an instruction, a dataset and evaluation tasks. Both forms should share a verifiable foundation.

Suppose a new paper changes the known limits of a method we use. The academy should help identify the related lessons, calculations, agent instructions and projects that need review. Updating knowledge then becomes a concrete task with consequences.

The academy should also preserve masters' experience: the signs they notice on site, their reasons for doubting a calculation, how they judge equipment condition and when they stop an experiment. Such experience becomes teaching material after its conditions and limitations have been examined and clearly described.

5 Why a real master is needed

A master sees the whole subject. He or she understands the required outcome, who needs it, its cost and how to establish that the work is complete. The master retains deep competence in their own craft and knows how to involve people with other specialisms.

The academy broadens that view. A lighting professional benefits from recognising when a question enters biology, requires an economic calculation or depends on operation. Deep specialist work can be shared between agents and relevant professionals. The master's understanding should support asking the question, assessing the answer and spotting a dangerous simplification.

The role has concrete responsibilities: define the goal and quality criteria; check inputs; establish each participant's authority; organise measurement and testing; decide within one's competence; and support the result through operation. Beginners approach this role through practice supervised by experienced people.

Professional judgment develops through engagement with reality. It includes recognising what is still unknown. A master should be able to stop, involve another specialist, require a repeated measurement and acknowledge that an attractive solution is not yet ready for use.

Agents help retain detail, return to unfinished actions, notice changes and document results. Human colleagues contribute additional expertise and work on site. For us, completion, reliability and durability test mastery, including the period after handover when the solution continues to operate without its author present.

6 How the model works in the Light Academy

Suppose a store needs new lighting. The master and customer first define the tasks: product visibility, working conditions, the quality of the space, energy use and maintenance. Any proposed effect on performance or behaviour requires its own measures and an appropriate evaluation method established in advance.

Work is then distributed. One agent gathers inputs and identifies gaps. Another reviews scientific evidence. A third helps calculate the lighting environment. A fourth compares controls and energy use. A fifth calculates ownership costs. A reviewing agent looks for incompatible assumptions, weak sources and errors in units. The master and relevant specialists examine the combined findings.

The academy should teach the proper scope of each conclusion. Illuminance measures one characteristic of the lighting environment. A claim about fewer staff errors requires observations of errors. A claim about increased sales requires sales analysis that accounts for other factors. Connections between these outcomes must be established.

A pilot compares the design with a real site. Baseline conditions are recorded, changes documented and a justified comparison chosen. Season, assortment, footfall, staffing and other important factors are considered when estimating the result. The estimated effect is accompanied by an account of uncertainty.

Operational experience then returns to the academy: deviations, repairs, complaints, successful choices and reasons for revision. It informs materials for people and instructions for agents. Learning receives feedback from a working installation.

7 Trust is built through evaluation

We propose assessing the combination of professional, agents and tools on a defined class of tasks. The person should demonstrate task formulation, understanding of limits and the ability to check outcomes. Agents should complete tasks in which answer accuracy and action quality can be established.

Evaluation should include difficult cases: incomplete inputs, an outdated method, conflicting sources, an implausibly favourable result or a task beyond the system's competence. A good system should recognise when it needs more information or a specialist.

Agreement among several agents has limited evidential value. Shared models, sources and instructions can reproduce the same error. Significant conclusions therefore need checking against primary data, calculations, measurements and, where necessary, independent expertise.

For operational qualification we propose recording model and tool versions, conditions of use and a review date. Material system updates should trigger the relevant evaluations again. This internal competence assessment needs an explicit status and scope.

The academy itself should also be assessable. Its quality can be judged through material reliability, correction speed, graduates' ability to solve real problems and the quality of equipment support. Lesson attendance and the number of certificates cover only part of education.

8 Who is developing related approaches

Our position has intellectual predecessors and practical counterparts. They share individual principles with our model and differ in purpose, scale and maturity.

Douglas Engelbart. In Augmenting Human Intellect (1962), he treated a trained person, tools, language and methods as a whole system. Human capability develops through coordinated changes to these components. For our academy, this provides an important foundation: tools, user preparation and work organisation should evolve together. [1]

Google Co-Scientist. In May 2026, Google described a Gemini-based system whose specialised agents propose, critique, rank and refine scientific hypotheses for researchers. On the same page, MIT biologist Ritu Raman explains how it helps her examine literature, structure ideas and identify expertise to seek from colleagues. This closely illustrates a scientist, agents and a human research team working together. Published applications concern specific tasks and require domain-level validation. [2]

Anthropic Research. Its 2025 engineering account describes a lead research agent coordinating specialised agents across different lines of inquiry. The authors discuss task definition, tools, evaluation and coordination, alongside computational costs and limitations for tightly coupled work. This supports matching the design of an agent cloud to the work involved. [3]

Researchers at Harvard Business School and Wharton with the P&G team. The Cybernetic Teammate, published in June 2026, reports an experiment with 791 Procter & Gamble professionals. In product-innovation tasks, AI users combined technical and commercial perspectives more effectively; individuals using AI matched teams without AI. The authors also identify the continued value of human judgment in selecting ideas. Authors include Ethan Mollick and Karim Lakhani. The evidence supports combined competence under the studied conditions; other professions require their own evaluation. [4]

The US Department of Energy and Genesis Mission. The official programme connects scientific data, computing, experimental facilities and AI, linking research with energy and national security. Scientific infrastructure and intelligent systems work together at its core. This illustrates an institutional move in a related direction. Programme goals should be distinguished from achieved results. [5, 6]

These examples show that combining professional human expertise with distributed intelligence is already a serious research and development agenda. Our task is to apply this logic consistently to an industry academy, the preparation of masters and the support of real installations.

9 Our founding position

The knowledge domains we work across are too extensive and change too rapidly for complete individual coverage. We therefore focus on the continuous development of core intelligence and specialised agents. The academy should provide verifiable knowledge, methods and feedback from practice.

For people, it provides orientation and a path to mastery. It helps them see the whole subject, ask precise questions, manage agents, involve the right specialists and verify outcomes. Their own professional preparation, ability to act and responsibility for decisions remain essential.

We want masters to work with a breadth of knowledge that once required a large organisation and draw on people wherever their experience and participation are needed. Their attention should reach the installed lighting, equipment condition, human comfort and the project's life after handover.

This is the work for which we are building the Light Academy and our educational resources.

Text for academy entrance pages

The knowledge domain covered by this academy is extensive and constantly changing. We focus on the continuous development of core intelligence and specialised agents. The academy helps people understand the subject, develop mastery, set tasks, manage agents and verify outcomes. The professional remains central: he or she connects knowledge, people and tools, makes decisions and takes responsibility for the quality of what happens in practice.

Sources

Sources checked on 26 September 2026. The estimate of 350–600 papers per day belongs to the authors and requires separate bibliometric confirmation.

[1] Douglas C. Engelbart. Augmenting Human Intellect: A Conceptual Framework. Stanford Research Institute, 1962. https://dougengelbart.org/content/view/138/

[2] Google DeepMind. Co-Scientist: A multi-agent AI partner to accelerate research. 19 May 2026. https://deepmind.google/blog/co-scientist-a-multi-agent-ai-partner-to-accelerate-research/

[3] Anthropic. How we built our multi-agent research system. 13 June 2025. https://www.anthropic.com/engineering/multi-agent-research-system

[4] Fabrizio Dell’Acqua et al. The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork. Organization Science. Published online 12 June 2026. DOI: 10.1287/orsc.2025.20702. https://pubsonline.informs.org/doi/10.1287/orsc.2025.20702

[5] U.S. Department of Energy. The American Science and Security Platform. https://www.energy.gov/undersecretaryforscience/genesis-mission/american-science-and-security-platform

[6] U.S. Department of Energy. Genesis Mission. https://genesis.energy.gov/