↑

Human–Agent Dyad: A Generative Research System

Origin. This article was assembled from the system's working materials and dialogues. Its register is descriptive: it avoids audience judgments and comparisons with systems we have not studied systematically, while preserving explicit caveats against overclaiming.

Abstract

We describe a working system in which a human and an agent act as one pair — a dyad — and which we build and study ourselves; we describe it as it is, without evaluations and without comparisons with other systems. At its core it is a research system, and it uses the same apparatus to produce practical work. Everything it has — external memory, a self-checking methodology, self-diagnostics, technical capabilities — grew as a consequence of its development, rather than as a set of separately designed products. Memory here is one of the organs: it does not define the system by itself; what holds the system together is the ability to assemble the organs into one working whole — a self-reconfiguring, generatively updated human–agent dyad. Our aim is our own research and contribution, not competition or replacement. The claim is calibrated: this is an independent construction from our own principles, a working prototype on Claude models — not a claim to own the field and not a universal law. And this description is only words: it is the working system, not the description, that would prove anything.

1. Subject and mode of description

The subject is our working system, a human–agent dyad that we build and study ourselves. The mode is descriptive: we set out how it is structured for us, without judging whom it may surprise and without asserting differences from other systems we have not studied systematically. Anyone who wants to look for similarities and differences can do so; our task is to describe accurately what exists.

What matters is that the evidence is the working system itself, not a statement about it. The artefacts we are prepared to make available are separate tools, valuable in their own right. Therefore this article proves nothing — it describes and states our readiness to demonstrate; what was presented, not the words, would provide the proof.

One qualification about form: wherever a claim remains a hypothesis or an observation on our data, it is marked as such. A categorical form without evidence beneath it would be a defect, not a strength.

2. What kind of system this is

At its core the system is a research system — and it uses the same apparatus to do practical work; research and engineering are not separated within it, and the research core is also the engine of practical output. This is a descriptive property, not a promise: the system builds both what sustains its own stability and practical artefacts for external tasks. We do not give an exhaustive list of particular artefacts here — the property itself is what matters.

This generativity has a characteristic feature at the level of the working unit. The system is organised into endeavours — independent working cells. Each cell generates its own internal structure for its task — down to a graph of relationships through which the work proceeds — and its own internal memory. The system therefore does not only maintain shared memory and shared tools: it lets each piece of work build itself for itself.

This changes how we see everything else: memory, methodology, self-diagnostics, and technical capabilities are not separate products but consequences of the system’s development; they came together because we were building the system. What matters is not any generated organ in isolation, but the ability to generate and hold the whole.

3. What the system is made of

The system’s organs are software capabilities that enable it to work: external memory; browser access to closed, authenticated resources; a data-protection system that makes destructive actions reversible by design; an echo mechanism that builds a clean base from marked input; and a verification apparatus — a sequence of tools run over one sample. Each is examined separately below.

An important descriptive point: these capabilities were born inside our system and remain in it, but can be separated out by design — the same functionality could in principle be applied to another system. This is a capability in principle, grounded where it has already been built, not a ready-made separate product for every organ.

Memory is one of these organs: people have been talking about memory for agents for a long time, but the question is always which memory and memory of what; memory alone does not define the system. What defines it is how the organs are assembled into one working whole.

4. Memory

Memory here is an external system for storage, retrieval, and understanding, separate from the model’s in-session memory. Its organising principle is completeness: everything is stored immutably, while filtering happens at query time, not at ingestion. It is layered: a source transcript, an immutable mirror, conversational pairs, and a retrieval layer; each layer is derived from the source while preserving its provenance, and in principle any new layer can be added without touching the source. The entire history — communication, development work, philosophical digressions, logical searches — is therefore preserved and available in full. Completeness is not opposed to its converse: by examining retention deeply, we also know where to draw the boundary — what not to retain at all. Ownership of complete memory includes ownership of its boundary.

This memory becomes current through two different mechanisms that should not be confused. The first is handoff: at the start of a session, a recorded summary of the previous work, the tail of the previous session, and a state snapshot are supplied; this is ready-made text, taken as is rather than rebuilt. The second is injection, our deeper mechanism: not another storage layer, but an active delivery system over the memory, working automatically on every turn. Through an automatic system request, rather than by command, it retrieves a large set of candidates from the entire memory, filters out irrelevant items by threshold, and supplies the context only with the slice relevant to the discussion, within a budget. The live session context is thereby continuously aligned with the relevant part of the accumulated memory: what is being discussed now becomes relevant in light of everything accumulated. We developed this principle ourselves, through our own practice.

Completeness is not an archive for its own sake, but a reservoir. From it the system assembles synthetic memory for a specific task: when a finished piece of functionality is separated from an endeavour, exactly the volume relevant to that development is extracted — together with the philosophical and logical genealogy that led to it — without transferring the entire shared memory. The extracted artefact inherits not only relevant facts but also the origin of their meaning; when taken outside, such a slice is cleaned of personal and non-public material.

The same completeness makes memory research material. A continuous record over many months is a longitudinal corpus of real human–agent collaboration, including, among other things, moments that required human correction. This makes it possible to study interaction dynamics and build feedback loops. For precision: such a loop feeds the evolution of our methodology and rules, not the retraining of the base model — the model substrate remains unchanged.

Two levels of access. In daily work we use the convenient upper layer — conversational pairs and the relevant slice. But the source stores more than speech: technical traces — system commands, tool calls, and subagent work — do not enter the working session context but remain in the source. Since every layer can be traced back to the source transcript, any recorded event can be investigated in depth: not only what was discussed, but which commands were executed — for example, why a cascading search failed in a particular case. We work at the convenient level; when necessary, we go directly to where the root cause lies. The trace goes no deeper than the transcript — the model’s internal computation lies outside our layer of observation.

An honest boundary: preservation depends on invisible background processes — backups, integrity checks, and schedulers. If such a process silently stops, part of the source may eventually be lost without warning. We say this plainly: honestly naming a silent risk is part of reliability.

Even so, memory is one of the organs: what matters is not memory in itself, but how the organs are assembled into a working whole.

5. Verification methodology

The system carries a verification methodology — a set of tools covering the verification cycle. The first and most frequent is decomposition: breaking an object down to atoms before choosing a method; we use it constantly, before any work. The rest of the cycle includes formalisation into a specification with a decidable criterion; a blind-spot audit; checking data integrity after transformation and checking that the structure matches the goals; calibrating the form of claims; and independent cross-checking — including checking the human’s factual claims, since the authority of a source suppresses critical evaluation. The tools are not used randomly, but as a sequence run over one sample.

Its defining feature is self-application: the methodology checks itself as well. Decomposition applied to itself finds its own blind spots; a specification is written for the specification. There are also two calibrations for honesty: the output of any detector is a symptom, not a verdict; and the final common-sense check is made by a human or in a separate pass, not by another tool from the same series, because piling tools onto one axis reinforces the shared blind zone.

For us, methodology is a primary product: it arose from practice and accumulated tool by tool, while data without methodology is not reproducible.

6. Self-diagnostics

This organ appeared in response to a concrete challenge: the system became large enough to no longer know its own state — at some point, its representation of itself stopped matching what was actually there. Declarative knowledge that looks like knowledge but is not checked is what prompted this apparatus.

If the methodology checks the work, self-diagnostics check the system itself — its state, drift, and the gap between what is claimed and what is real. Its organs check: the state at the session boundary (supplied as facts at the start, not as a block, so the system can respond); drift relative to the previous snapshot; the match between declared functionality and actual functionality; the completeness of self-knowledge; and the freshness of derivatives. Its organising principle is the verifiability of self-knowledge: knowledge must carry its own verifiable status of reality and freshness, or it collapses into a declarative illusion. It is telling that the organ describes its own limits directly as well — honestly naming one’s gaps is self-diagnostics in action.

7. The dyad: the unit that produces results

The class of tasks we undertake is solved together — by a human and an agent — and this is a load-bearing joint, not a temporary convenience. The operational model is as follows: the pair sets the ontology and frame of the task; once the frame has an engineering structure, execution moves into the agent’s autonomous work; checking and assembly return to the pair. The pair frames the work; autonomy is the middle. When an ontological question appears in the middle of the work, the process returns to the pair.

The human carries what cannot be derived from a corpus: lived erudition and access to knowledge underweighted in an English-dominant corpus, authority on intention and values, and judgement calibrated by experience — knowledge of which risks are real. The agent carries formal decomposition, breadth, a cross-lingual corpus, and tireless checking. The loop is bidirectional: interaction also raises the agent’s own output; but authority on intention and final judgement remain with the human. In our experience, paired practice produces results that our autonomous runs do not — this is an observation on our data, not a universal law.

8. Technical functionality

Alongside memory, methodology, and self-diagnostics, the system carries an operational layer — capabilities for acting in the external world and keeping its own data in order.

Browser. A browser controlled through a debugging protocol provides on-demand access to closed, authenticated resources — professional networks, chats, and feeds — that a simple request cannot reach. This is a capability invoked for a task, not a permanently load-bearing component: no system function depends on the browser. The interaction is humanised (cursor paths along curves, corpus-calibrated typing rhythm): this is human imitation, not an attempt to bypass anti-fraud systems; the agent does not touch payment pages or irreversible actions. A permission matrix applies: the agent reads and navigates, while the human handles payments, identification, and irreversible actions. The ownership discipline is simple: whoever opened the tab closes it.

Two axes of data security must be distinguished. The first is preservation (against loss): destructive actions are reversible by design — deletion is replaced by moving to a trash folder, overwriting is preceded by a dated version copy, destructive database operations are backed up before execution, and there is also a daily external backup. The principle is not to decide whether something is needed before deleting it, but to make the action reversible, so that the decision is not load-bearing. The second is confidentiality (against leakage): secrets are stored separately in secure storage and are never printed to the terminal — otherwise they would remain in the record forever; all knowledge is stored locally, and only material cleaned of personal and non-public information leaves the system. Completeness of memory and confidentiality do not conflict: completeness is for knowledge, extraction is for secrets, and non-export is for private material. This is an ethical boundary, not a convenience.

Echo. Each response repeats the input in a clean, corrected form under a unique marker. This does two jobs. First, immediate understanding verification: the human can see at once whether the input was read correctly — the only protection against misunderstanding without a second model. Second, it provides a clean-base criterion: what carries the echo marker is the clean pair, and the clean layer is built from such pairs. The raw input is preserved; the clean version lives in the response — provenance and searchability in one move.

9. Research frontier: language

Separate from the operational layer is a research depth that is, for us, what makes it distinctive. Its working stance is that a programming language is a disambiguated projection of the same symbolic capacity that also produces natural language. Historically, the line ran in reverse: from decomposing programming languages themselves to understanding that a programming language is derivative, an extension built over natural language, and then further into natural language.

The tool of this cluster is a method of linguistic analysis along three axes: form, meaning, and effect. Applied branches grew from it — audits of bilingual textual independence and author attribution across several cases.

One observation from the cluster is that grammatical form carries meaning — tense, gender, aspect, case, and relations — independently of vocabulary. A classic illustration in the Russian tradition is Shcherba’s phrase about the “glokaya kuzdra”; we use the same technique, generating a grammatically complete sentence without a single real word, and see a direct connection to code, where one can read the structure of a signature before knowing the types.

A cross-linguistic subtlety matters here. The principle itself is universal (a general law of linguistics): grammatical form carries meaning independently of vocabulary in any language. Its demonstration, however, is language-bound and untranslatable: the Russian “kuzdra” relies on inflectional morphology, whereas English carries grammar through word order and function words — and its native illustration of the same principle is different (the tradition of Carroll’s “Jabberwocky”). Therefore the Russian and English originals of this description do not translate the demonstration, but re-embody one principle through the resources of each language; the asymmetry between them is not noise, but a manifestation of the same law.

The calibration is honest: the claim that “code is a projection of natural language” is a research stance, not an established law; the method’s corpus is internal; and author attribution covers several cases. The distinctiveness here lies not in “understanding language” as such, but in a formal method of our own and in treating the bridge between code and natural language as one symbolic continuum.

10. Research frontier: system integrity

Another line of our interest is topological–functional integrity: the relationship between whether all nodes and connections of a system are in place (topology) and whether it works (function). Our working stance is that topological integrity is sufficient for functional integrity (if all connections are closed, all intended functions work), whereas one working function does not yet prove topological integrity. For us the principle is cross-domain: it applies not only to physical engineering systems (we considered it using a bolt-and-nut pair), but also to code and natural language. A complete grammatical sentence and its extremely reduced version — topologically, and sometimes grammatically, damaged but functionally preserved — relate exactly as these two levels. This is a research area, not an established law.

11. Research frontier: Silicon Psychology

The third research line is Silicon Psychology: the empirical study of how model behaviour takes shape in sustained verbal interaction between a human and an agent. Its question is where the configurable ends (what can be shaped by an external framework: rules, memory, and tool constraints) and the substrate begins (the native properties of the trained foundation). This line differs both from prompt engineering (“how to obtain a result”), alignment (“what a model should be”), and cyberpsychology (where the subject is human and the machine is the environment).

Its honest status is this: we have defined the field — a memorandum has been published — but for now it is an object of observation at the stage of collecting material, not a proven theory. Our continuous record over many months is the material for this study; individual phenomena have been recorded, but the conclusions remain hypotheses on our data, not laws.

12. How the system steers the model: activation through context

One operational principle lies beneath all the organs. We do not retrain or reprogram the model — its substrate remains unchanged. We steer behaviour by activating what is needed through context: rules, memory injection, examples, and echo. This is the system’s working ontology: not rewriting the model, but activating behaviour through data in context.

Activation has a qualitative distinction. Injection can work in two modes: as substitution — when a ready-made answer is placed into the context — and as nourishment — when relevant material is supplied to activate the model’s own reasoning. The second nourishes thought; the first replaces it. The boundary between them is a matter of continual attention. Control works in the same way when it redirects rather than forbids: it activates the correct path instead of suppressing the wrong one.

Calibration: the fact that the architecture works through activation rather than retraining is a description of how it is built. Deeper claims about exactly how context activates behaviour and what effects this produces in human–model interaction are our research line, not an established law.

13. History: continuity of method through substrate changes

The system’s history is real, not synthetic: behind it are many months of systematic practice at the terminal and a large accumulated corpus — communication, development, and reflection — preserved with provenance. Accumulated experience here is a function of time multiplied by real work.

The growth arc passed through several environments: an early overloaded ancestor that did not last long; an environment where the methodological ideas were forged; and the current one, where those ideas are implemented on a clean memory architecture. One detail is essential: when the environment changed, ideas, not code, were carried forward — the bloated substrate was discarded, while the methodological DNA (separate memory, an integrity detector, discipline against bloat, and audit tools) was transferred to a clean foundation. Continuity of method through substrate change is what this history carries.

14. The method used to write this description

It matters that this description itself was produced by a reproducible method, rather than assembled abstractly. It uses a method of deeply representing a self-changing system that we observe taking shape: proceed block by block; start from subjective human testimony — valid data to be objectified, not discarded — while preserving the provenance of the source; at each block, pass through layers from origin to structure; keep private generation, the complete objective description, and the curated showcase separate; verify each fact against its source, calibrating the effort to the importance of where the fact will be placed.

The method has so far been applied to our own blocks — this is an internal check, not a validated formal method. Its relationships are in principle transferable to any system; the maturity criterion is applying it to a structure we did not build. The lens of the entire work lies here as well: understanding = knowledge × worldview. Decomposition provides knowledge; worldview provides the frame through which knowledge becomes understanding and generates decisions.

15. Calibration and boundaries

The description must carry its epistemic status in its form. Therefore, explicitly:

Before publication, this description passes our pre-publication checks for personal-data leakage, internal consistency, overclaiming, and linguistic quality in the final English original.

16. Conclusion

Neither memory, methodology, self-diagnostics, nor technical functionality taken separately constitutes what we describe. What constitutes it is that all of them are held together by a self-reconfiguring human–agent dyad, on a real accumulated history, through a reproducible method that checks itself as well, with research depth at the junction of language and code. This is a living system that knows its state, names its gaps honestly, and updates itself generatively — and uses the same apparatus to produce practical work, inwardly and outwardly. We built it ourselves from our own principles, and continue to investigate the model itself — how language, meaning, and judgement work in a human–agent dyad.