# Agentic Institutions: An Introduction

*How useful lessons can survive the next AI session*

**Luuk Joseph Soons · CounterProof Research**  
**Draft v0.1 · 28 October 2026**

*About ten minutes to read.*

## 1. The correction is saved. Why is the mistake back?

Imagine a research team discovers that one of its estimates used the wrong denominator. The apparent success rate is unreliable. Someone withdraws it, writes a careful explanation and saves the correction in the team's library.

A week later, an AI assistant preparing a new report retrieves the original summary and repeats the estimate. The correction still exists. The archive has preserved both the mistake and the reason to reject it.

This is a hypothetical example. It poses a practical question: **what must happen for a lesson learned in one session to improve the work of the next?**

Saving the explanation is a beginning. A later participant must find the relevant record, receive enough of it to understand its bearing on the task, and use it where it matters. Someone must also decide what the correction changes: the estimate, the report, the calculation procedure, or all three.

Yet “follow the latest note” would be a poor solution. Suppose the correction itself uses the wrong comparison. Or it applies to last month's data, while today's task concerns a different sample. A capable successor should examine the grounds and scope of the correction, rather than obey its date or its author's title.

Agentic Institutions is our proposed research programme for these questions. It asks how arrangements of people, AI systems, tools and records can sustain useful work across successive sessions while remaining open to criticism. The paper offers a framework and a study proposal. It has not demonstrated that its proposed procedure improves practice.

## 2. An institution can begin with a small team

The word *institution* can suggest a university, a government department or an imposing building. Here it refers to enduring working arrangements. A small team already has some: who assigns tasks, which sources it uses, who may change a shared instruction, and what happens when someone disputes a decision.

An agentic institution brings AI participants into those arrangements. Its library, working rules and tools can remain available after an individual session ends. Continuity need not reside in one enduring assistant. It can reside partly in what people preserve and how they make it available to successors.

This does not require a new legal entity or imply an autonomous company. Giving an assistant a job title does not make it accountable. People still determine the purposes pursued, the access granted and the consequences accepted.

The underlying problem is older than AI. Research on organisational memory asks how retained experience affects later decisions. Work on organisational learning examines how routines improve or entrench mistakes. Distributed cognition examines work done across people, tools and documents. The present programme draws on these traditions.

AI changes some of the conditions. A fresh session can begin in seconds, with instructions and source material assembled automatically. Models, tools and access permissions can change while the documents keep the same names. We need to examine what these arrangements pass on, and what their successors do with it.

An arrangement can be persistent, distinctive and wrong. What matters is whether later work responds better to relevant evidence.

## 3. Follow the lesson into a decision

Return to the withdrawn estimate. Several different failures could produce the same incorrect sentence in the new report.

Perhaps the correction never entered the searchable library. Perhaps the search missed it. Perhaps a shortened handover omitted the explanation. Perhaps the assistant received it but applied it to the wrong estimate. Or the discrepancy was recognised and nobody had authority to change the report.

Those failures need different remedies. More storage would not repair an inappropriate decision about authority. A clearer note would not help if the retrieval process never selected it.

The paper therefore separates keeping evidence, selecting and delivering it, interpreting its scope, deciding what to accept, acting on that decision, and carrying the result into later work. These are questions to investigate, not a rigid sequence that every task must follow. A record of delivery does not reveal a model's internal understanding; we examine its expressed account and subsequent actions.

The proposed correction record is meant to help a successor answer ordinary questions: what was claimed, what supports it, what changed, who accepted the change, and where does it apply? It should also preserve important objections and identify later decisions that depend on the claim.

For a simple case, a dated annotation and a link may be enough. An elaborate record could cost more than it saves. Nor does a field marked “accepted” establish truth. It records a decision that remains open to examination.

The paper includes a narrow illustration from CounterProof's own documents: an earlier formulation appears alongside records withdrawing its measurement basis. This establishes their coexistence in an inspected snapshot. It does not show that a later agent encountered the old wording, relied on it or repeated the error.

## 4. Why not use better memory and better models?

The strongest objection is also the most useful: much of this sounds like knowledge management. Keep dated sources, correct obsolete summaries, link revisions and assign responsibility. Why introduce another programme?

Those practices are serious starting points. If they solve the problem at acceptable cost, a more elaborate procedure has no claim on the team's time. The proposed research should identify where established practices suffice and where a specific addition helps. Renaming familiar work is not evidence of an advance.

Retrieval-augmented generation, or RAG, helps supply an assistant with material selected from a collection. Memory systems can preserve and organise earlier work. Such systems may also include correction handling and approval. The question is how well a particular arrangement supports the later decision, including when its retained guidance is mistaken or no longer applies. Retrieving both conflicting notes leaves work to do; it does not follow that an additional formal procedure is always needed to do it.

Better models may also reduce the problem. A stronger successor might recognise the denominator error immediately and need little help from a correction record. That would be a reason to simplify the procedure. Conversely, an improvement after replacing a model would not, by itself, show that the institution had learned from its history.

There is no commitment here to preserving machinery once it becomes unnecessary. The research question concerns useful continuity. Any particular answer must earn its place against competent alternatives.

## 5. More than a record of who approved what

A good audit trail lets someone reconstruct a decision. An institution must also decide who can challenge it and what happens when the challenge is justified.

Suppose the recurring mistake comes from a standing instruction that tells every assistant to use the same unsuitable calculation. Correcting each report leaves the cause in place. Someone needs authority to revise the instruction, explain the change and withdraw its obsolete guidance from current use.

The proposed arrangement distinguishes proposing a change, authorising it and examining the reasons for that authorisation. A small team may combine these roles, but should make the overlap visible. For consequential disputes, additional scrutiny may be needed from someone who did not make the original decision. The founder's preferred rule must be challengeable too.

Human approval is not a guarantee. A tired or interested reviewer can accept a poor explanation. The proposal asks for visible responsibility and reasons that can be questioned, while allowing routine actions to be delegated within stated limits. If every minor task becomes a meeting, the arrangement may be making the work worse.

Nor does the institution need several models simply to deserve the name. Different providers can still share assumptions, sources and instructions. One model family may be used alongside effective external checks. Additional reviewers should be justified by what they contribute, with their costs included.

Finally, keeping a record need not mean repeatedly presenting it as guidance. An old claim can remain available as history while being retired from current instructions. The purpose is useful work, not the largest possible archive or the greatest number of approvals.

## 6. A first test that could favour the simpler approach

The first proposed study is deliberately narrower than the whole institutional ambition. It would hold the model and its operating setup fixed, then examine decisions across repeated handovers. Software-maintenance cases would be the primary setting, with research-account cases treated separately.

Each case would introduce changes in evidence: a valid correction, a mistaken objection, a partial retraction, or a result that applies only to a particular software revision. Some disputes would remain unresolved. Making every newest note correct would reward obedience rather than judgement.

The main comparison has two approaches:

- **Ordinary repair and checking.** Retrieve dated records, annotate corrections, link old claims to their replacements, and check sources, conflicts and where each claim applies.
- **The proposed procedure.** Use the same resources and ordinary checks, adding the correction record and a specified procedure for accepting, challenging and revising guidance.

Both would receive the same evidence and substantive history. A judgement or qualification in the structured record must also appear explicitly in the ordinary notes. Otherwise we might mistake the benefit of extra human explanation for a benefit of the procedure.

Resource allowances would be matched. Actual model use, elapsed time and human work would also be counted. A later exploratory stage could let each approach develop its own records. Because those histories would then differ, it would answer a different question.

The main question is whether later decisions use relevant corrections appropriately, preserving their limits and uncertainty. The study would also count repeated errors, mistaken reversals of sound conclusions, uncritical obedience and human effort. More corrections would not automatically mean better performance.

Before the main test, the team must define worthwhile improvement, acceptable harm and cost, and when uncertainty makes a result inconclusive. If ordinary repair performs just as well within those criteria, the extra procedure should be dropped or redesigned. We must also distinguish a procedure that was not implemented properly from one that received a fair test and added nothing.

**No such experiment has been run.** The detailed protocol, thresholds and staffing remain unfinished. Separate roles for authorising changes, preparing reference answers and assessing outcomes must be filled before the main test. Replacing models across handovers is a future extension, not a result of this first design.

## 7. Keeping inquiry alive

The longer ambition extends beyond avoiding repeated mistakes. An institution might help people discover useful connections, develop ideas and pursue questions that outlast a particular project or assistant. Preserving corrections would support that work, but a successful handover test would not demonstrate creativity or the quality of an entire institution.

The underlying question also survives a negative result. If the proposed record adds no value, people will still need to decide what to preserve, what to question and when to retire inherited guidance. Learning that a simpler arrangement works better would itself be useful progress.

This programme sits beside two related inquiries. Agentic Stemmatics examines textual transmission and dependence, including what model agreement can support. Agentic Representation examines how selection and arrangement shape the account a reader receives. Institutions examines how people act on such accounts and carry, challenge or revise them over time.

Their relationships run both ways: institutions select sources, reports shape attention, and criticism can change a working rule. Success in one programme does not certify another. Each needs evidence appropriate to its own claims.

## Reading further

[Read the full research paper on this site →](https://research.counterproof.io/paper/agentic-institutions/)

The research paper is *Agentic Institutions: Organisational Memory, Correction and Continuity Across Successive AI Agents*, v0.3. Begin with §§1–2 for definitions, §§5–6 for the proposed procedure and example, §7 for the comparison, and §§9–11 for extensions and limits. Its DOI, [10.5281/zenodo.23727633](https://doi.org/10.5281/zenodo.23727633), is now registered; the paper was published on Zenodo on 28 October 2026. This identifier belongs to the research paper, not this introduction.

Related guides: [Agentic Representation: An Introduction](https://research.counterproof.io/intro/agentic-representation/) and [Agentic Stemmatics: An Introduction](https://research.counterproof.io/introduction-agentic-stemmatics.html).

*This introduction explains a conceptual proposal; it supplies no independent evidence for it and has not been peer reviewed. The author has a financial interest in CounterProof, whose practices could benefit from the proposed research. Prepared with Codex (OpenAI) assistance; responsibility for the argument and any publication remains with the author.*