# Agentic Epistemics: An Introduction
*How people and AI can build knowledge that remains open to correction*
**Luuk Joseph Soons · CounterProof Research**
**Introduction paper · 31 October 2026**
*About twelve minutes to read.*
## 1. What have we actually learned?
A laboratory tests a new temperature sensor. Its readings are accurate enough for the experiment. An AI assistant reviews the results and writes a summary: the sensor passed its tests. Months later, another assistant finds that summary and recommends the sensor for outdoor use.
There is a problem. The original tests covered a narrow range of temperatures in a stable room. They did not examine frost, summer heat or changes over time. The summary may contain no false sentence. The sensor may even work well outdoors. But the earlier tests do not establish that it will.
This is an invented example, not a result from our research. It illustrates a question that arises whenever work passes between people, machines and records: **what gives the next person good reason to rely on what they receive?**
In software, the same question appears when “passed these tests” becomes “ready for every intended use”. In research, it appears when a tentative explanation becomes a settled fact in a later account. An institution can preserve the words while losing the conditions that made them worth believing.
Agentic Epistemics is CounterProof Research's wider programme for examining such questions. It asks how AI can help us discover, judge and retain knowledge—and how we can notice when it leads us astray. This introduction explains that programme and its intellectual background. It does not claim that we have invented epistemology or demonstrated a new method.
## 2. An answer, a reason and an understanding
*Epistemology* studies knowledge: what it is, how we acquire it and what justifies our claims to possess it. Our use of *epistemics* directs attention to the practices through which people and institutions handle these questions. *Agentic* refers here to software systems performing connected tasks, using tools and passing work to others. The label does not establish consciousness or independent purposes.
It helps to separate three achievements.
**An answer becomes available.** A model proposes an explanation, writes some code or identifies a possible mistake. The contribution may be useful before anyone has established that it is correct. An imaginative conjecture can open a worthwhile investigation.
**There are grounds for accepting it.** Someone examines the evidence, checks the argument or runs a suitable test. What has been established depends on what was actually checked. Finding the same answer in several places may matter, but we need to know whether those places supply additional evidence or repeat one source.
**Someone understands what it means.** A reliable prediction need not explain the mechanism behind it. A proof can establish a mathematical result without showing that the result applies to the physical system we care about. Understanding may concern why something happens, how to investigate it or when a result can be used.
A claim can be true even when we lack good grounds for accepting it, and well-supported beliefs can turn out to be false. Whether to use a result is another judgement, shaped by the purpose and consequences of error.
These achievements can develop unevenly. We should welcome a useful conjecture without presenting it as an established result, and recognise a reliable result without claiming that every underlying mechanism is understood.
This also leaves room for a positive ambition. AI might help us formulate better questions, explore unfamiliar material, construct arguments and discover relationships that we would otherwise miss. The programme concerns how knowledge can grow, as well as how mistakes can be caught.
## 3. Why this is both old and new
Science has never depended on unaided individual minds. It relies on instruments, testimony, records and divisions of expertise. Most researchers cannot personally repeat every experiment or reconstruct every calculation on which their work depends. John Hardwig's work on epistemic dependence makes this difficulty explicit: reliance on others is part of how demanding scientific work becomes possible. [Further reading: Hardwig](#reading-hardwig).
Nor did automated discovery begin with chatbots. The 2009 report on the robot scientist Adam describes a system that generated and experimentally tested hypotheses in a bounded area of biology. Contemporary AI enters a longer history of tools helping to shape inquiry. [Further reading: King and colleagues](#reading-king).
What deserves examination now is how systems participate across several stages. An assistant may propose a result, assess it, write the report and supply the next assistant's background material. These connections can make work faster and criticism more accessible. They can also allow one mistaken assumption to travel through an apparently varied process.
A human can make the same mistake. For a claim with its meaning fixed, changing its author from a person to a model does not change whether it is true. What can change is how it was produced, what supports it and how cheaply it can be repeated. Those differences matter to warranted reliance.
There is already research on this. Sabina Leonelli examines how data acquire evidential significance as they move into new settings. Uwe Peters and Benjamin Chin-Yee study how AI summaries can broaden the conclusions of scientific texts. Our temperature-sensor example therefore illustrates an established family of problems. Our proposed contribution must be judged against that work, not against an imagined absence of it. [Further reading: Leonelli](#reading-leonelli) and [Peters and Chin-Yee](#reading-peters).
## 4. Learning from criticism and tradition
Popper directs attention to criticism: what could reveal that an explanation is mistaken, and whether we give that possibility a fair hearing. This is more demanding than asking for an adversarial-sounding response. A criticism must bear on the claim and its grounds. A disappointing result also needs investigation; it does not automatically tell us which assumption failed.
Kuhn directs attention to the traditions that make research possible. Shared examples, concepts and expectations allow people to work together and pursue difficult problems. Yet those same commitments can make alternatives hard to recognise. Popper and Kuhn disagreed about scientific practice; their differences should remain visible when we draw on them. [Further reading: criticism and tradition](#reading-criticism).
An AI institution faces a practical version of this tension. Without retained definitions and lessons, each new session repeats old work. Without the freedom to question them, the institution may become highly competent at preserving an error.
The aim is neither permanent suspicion nor unquestioning acceptance. It is to know which matters can reasonably be treated as settled for the present task, and what would justify reopening them. Institutions need continuity, but their rules must remain answerable to the work those rules are meant to serve.
## 5. Three research strands
The wider programme currently includes three connected inquiries. They approach different parts of the problem; success in one would not establish success in the others.
**[Agentic Stemmatics](/introduction-agentic-stemmatics.html) asks about origins, transmission and dependence.** Where did a claim come from? Why do outputs resemble one another? Does agreement supply additional support, or could it reflect shared material or shared mistakes? Different provider names do not by themselves establish independent errors. We therefore ask how each output was produced before deciding what its agreement adds.
**[Agentic Representation](/intro/agentic-representation/) asks what an account leads its reader to believe.** Are the conclusion, evidence and qualifications presented adequately for the question being answered? Accurate quotations can coexist with a misleading overall impression. Representation here concerns reports and accounts, rather than the internal vectors of a model.
**[Agentic Institutions](/intro/agentic-institutions/) asks what survives into later work.** Was a relevant lesson recorded? Did it reach the next participant? Did that participant use it appropriately? How can people revise the rules, resolve disagreement and retain responsibility across changing sessions?
Return to the sensor. Stemmatics could examine whether several favourable accounts descend from the same report. Representation could examine whether the summary preserves the test's limits. Institutions could examine why a successor does or does not take those limits into account. None of these inquiries alone establishes the sensor's outdoor accuracy.
Interpretability is a neighbouring investigation. Evidence about a model's internal computation may help answer questions that output records cannot settle. It does not automatically establish the truth of a claim about the outside world. Our [companion note on collation and interpretability](/note/collation-and-the-microscope.html) discusses how these approaches might inform one another.
## 6. What an institution should help us do
An agentic institution is a continuing arrangement of people, AI systems, tools, records and responsibilities. Its life is longer than any individual chat. A new assistant can inherit a library and a set of working rules, but inheritance alone does not establish learning.
Useful continuity requires judgement. Some earlier results remain adequate. Some need new checks because the task has changed. Some should be retired. A warning that was sensible for one system can become an unnecessary obstacle for another.
Responsibility also belongs at different points. The original investigator should explain what was assessed. Someone endorsing a summary should preserve the qualifications relevant to its intended use. A successor introducing a new purpose should examine the new conditions. Publishing a careful laboratory result does not make its author responsible for every unforeseeable later use.
This is why we study the passage of grounds as well as the retention of text. A larger archive may help, but a short, clear limitation notice may solve the problem more cheaply. Better models may make some proposed procedures unnecessary. The institution should be able to discover that and simplify its work.
The same discipline applies to CounterProof. Our research may suggest useful improvements to adversarial code review and software construction. Whether a particular procedure helps clients requires evidence about its results, limitations and cost. An attractive philosophy or a carefully maintained record cannot establish that benefit on its own.
## 7. Why music sometimes helps us explain this
Imagine a musician receiving a marked score. One instruction comes from the composer, another from an editor, and a third from a previous performer. The marks may be useful, but they do not all carry the same authority. Understanding their origins helps the musician decide what to preserve, question or change.
Now imagine a school passing that score to new generations of players. It preserves techniques and interpretations that would be difficult to rediscover alone. It may also preserve a mistake. A living tradition needs both memory and the possibility of reconsideration.
This is a teaching comparison, not evidence that AI models behave like musicians or that resemblance proves descent. It makes otherwise abstract questions tangible: where an account came from, how it is interpreted and what a successor inherits. Our essay [Why music belongs in our research](/essay/why-music-belongs-in-our-research.html) develops these connections and their limits.
## 8. Our starting commitments
CounterProof's wider project grows out of cypherpunk concerns with privacy, voluntary cooperation and limits on concentrated power. Those concerns lead us to ask who can participate in an institution, question its rules and choose how much to delegate. They are commitments we should explain and defend, not proof that any preferred architecture succeeds. [Further reading: Hughes](#reading-hughes).
Austrian influences add questions about purposes and knowledge. Mises's praxeology studies purposeful human action: people choosing means in pursuit of ends. Hayek asks how the knowledge dispersed among people becomes available for their decisions. Applied here, these questions concern who sets an AI system's purposes and what grounds make it a suitable means. They do not establish that software has subjective purposes of its own. [Further reading: Mises and Hayek](#reading-action).
These influences are not one harmonious theory. Mises's account of a priori knowledge and Popper's account of empirical testing differ. Our contingent claims about AI systems and institutional arrangements must face appropriate evidence, whatever our intellectual sympathies.
Clavestra's proposed federations belong to this longer-term ambition. They concern choices about monetary infrastructure and delegation; they have not been implemented or launched. Their possible merits require their own investigation. An institution can do serious research using existing services, and a reader can reject our preferred monetary arrangements while finding the epistemological questions worth pursuing.
## 9. Would more capable AI make this unnecessary?
It might make particular safeguards unnecessary. An assistant could become much better at preserving qualifications, checking arguments or recognising when a task exceeds the evidence. That would be progress. We should not preserve a procedure merely because we named it or invested in it.
The broader question would remain: what has the system established, and what justifies relying on that achievement here? A model could help answer that question too. Human involvement is not a guarantee of correctness; people can overlook evidence and misunderstand a convincing explanation.
Nor must we decide whether a model itself knows something before assessing every contribution it makes. A generated proof can be examined. A proposed explanation can guide an experiment. A useful objection can expose a mistake. The grounds for accepting each contribution depend on its character, rather than on treating human or machine authorship as an automatic verdict.
This leaves genuine questions open. Could AI help produce science beyond any individual's understanding? Which parts would need to be understood, by whom, for a community to use it responsibly? How would we distinguish an unfamiliar discovery from a convincing illusion? Existing debates about AI, understanding and scientific discovery are part of the programme's starting point. [Further reading: AI and science](#reading-ai-science).
## 10. What would count as progress?
For this programme, progress would include making a question clearer, exposing a mistaken assumption or finding a better way to carry justified conclusions into later work. It would also include discovering that an ordinary practice already solves a problem and an elaborate proposal adds nothing.
Some questions call for philosophical argument. Others call for historical investigation, close examination of records or controlled comparison. We should choose the method for the question, and state what its answer can support. A thought experiment can reveal a distinction; it cannot tell us how often a deployed system fails.
An introduction has a simpler job. It gives readers a way into the questions, explains how they connect and points to work that develops them. That is the purpose of this paper. The linked research programmes contain proposals whose effectiveness remains to be established.
The ambition is to help people and AI discover more while keeping their conclusions open to serious criticism. An institution worthy of that ambition must be able to improve its answers—and to change the questions it asks.
## Reading further
For an accessible practical companion, read [When does an AI answer become knowledge?](/essay/when-does-an-ai-answer-become-knowledge.html). The [introduction papers](/intro/) explain the individual strands; the [research papers](/paper/) provide their fuller arguments and proposed studies.
**Knowledge through others.** John Hardwig, [“Epistemic Dependence”](https://doi.org/10.2307/2026523) (1985), examines reliance on expertise when no individual can reconstruct all the relevant grounds.
**An earlier example of automated inquiry.** Ross D. King and colleagues, [“The Automation of Science”](https://doi.org/10.1126/science.1165620) (2009). The [institutional record and original abstract](https://research.manchester.ac.uk/en/publications/the-automation-of-science/) describe the bounded Adam system; our account here relies on that abstract, not a fresh evaluation of its experiments.
**Evidence in changing settings.** Sabina Leonelli, [“What Counts as Scientific Data? A Relational Framework”](https://pmc.ncbi.nlm.nih.gov/articles/PMC4747116/) (2015), examines the movement and use of data as evidence.
**The scope of a scientific summary.** Uwe Peters and Benjamin Chin-Yee, [“Generalization bias in large language model summarization of scientific research”](https://pmc.ncbi.nlm.nih.gov/articles/PMC12042776/) (2025). Their comparison concerns whether summaries broaden source conclusions; it does not independently establish that those source conclusions are true.
**Criticism and scientific traditions.** Karl Popper, *The Logic of Scientific Discovery*, especially §§3 and 22; Thomas Kuhn, *The Structure of Scientific Revolutions*, especially the 1969 Postscript. Their direct exchange appears in Lakatos and Musgrave's *Criticism and the Growth of Knowledge* (1970): [Kuhn's contribution](https://homes.luddy.indiana.edu/jbollen/I501F13/readings/kuhn_en.pdf) and Popper's “Normal Science and its Dangers”.
**Cypherpunk commitments.** Eric Hughes, [*A Cypherpunk's Manifesto*](https://www.activism.net/cypherpunk/manifesto.html) (1993).
**Purposes and the acquisition of knowledge.** Ludwig von Mises, [*Human Action*, chapter I](https://www.econlib.org/library/Mises/HmA/msHmA.html?chapter_num=4), and F. A. Hayek, [“Economics and Knowledge”](https://www.econlib.org/book-chapters/chapter-hayek-economics-and-knowledge/) (1937). Their arguments are sources for questions here, not empirical evidence for a proposed AI system.
**AI, understanding and discovery.** Kyle Cranmer and colleagues, [“AI for science: reframing AI's role in discovery”](https://doi.org/10.1093/rssdat/udag003) (2026); Emily Sullivan, [“Understanding from Machine Learning Models”](https://doi.org/10.1093/bjps/axz035) (2022); Tim Räz and Claus Beisbart, [“The Importance of Understanding Deep Learning”](https://doi.org/10.1007/s10670-022-00605-y) (2024). These works provide existing arguments and disagreements with which our programme must engage.
---
*Status and interests.* This is an introductory synthesis and guide to CounterProof Research's programme. It reports no new experiment, demonstrated method effectiveness or independently established novelty. It has no separate Zenodo deposit or DOI. Luuk Joseph Soons has a commercial interest in CounterProof's adversarial code-review practice. Prepared with Codex (OpenAI) assistance; responsibility for the text and its publication remains with the human author.