Intra-action: why “the human who uses AI” is a useful fiction — and a dead end

Prologue

We still say, in 2026, that the human interacts with artificial intelligence. Two stable entities — you on one side, the model on the other — come into contact, exchange, and produce an answer. This image is pedagogical. It is also ontologically false, in the sense that it hides what actually happens when an agentic system runs.

What I develop here under the name Agentic Realism, Response-able AI rests on a conceptual shift borrowed from the physicist and philosopher Karen Barad: moving from interaction to intra-action. This is not jargon to impress on LinkedIn. It is a design criterion: if you don’t take it seriously, you optimize chatbots; if you do take it seriously, you design apparatuses in which humans, agents, files and constraints co-emerge in every session.

I. Interaction: the myth of the two billiard balls

The classical picture of interaction assumes relata — “things” — that exist before the relation. Billiard ball A hits billiard ball B. The user opens an application; the application responds. The sovereign human subject commands the passive tool.

In AI productivity, the soft version of this myth fits in one sentence: “I use ChatGPT for my project.” Beneath that sentence sleep three assumptions:

1. Me (the user) am the same from one session to the next.
2. The model is a stable, generic capability.
3. The project is a real, already-defined object that the tool reflects or executes.

Barad shows, in another discipline (quantum physics, epistemology, science & technology studies), that this picture fails as soon as you describe how the real is produced in material and discursive practices. She proposes agential realism: agency — the capacity to act, to make a difference — is not a property an individual possesses; it is enacted in precise configurations of apparatuses, measurements, boundaries and language.

Minimal translation for us, the builders: useful intelligence is not in the model. Nor is it “in” the human. It emerges in the complete arrangement.

II. Intra-action: entities are born in the relation

The term intra-action (not inter-action) stresses the intra-: inside of, within. The “entities” do not precede the relation; they are distinguished through the relation, in a single gesture Barad sometimes calls cut together/apart.

It is not that independent entities enter into interaction; it is that, in intra-action, what counts as an entity is actualized by the apparatus.

A trivial but lived example

You open a session with an assistant, you load three files, you forget the fourth, you’re tired, the cron ran last night and left a partial summary in a folder, the system prompt forbids certain tools. You ask: “Where does project X stand?”

What happens is not: a stable user queries a stable model about an already-defined project X.

What happens is: a singular intra-action in which:

• “You” are co-constituted by your fatigue, your file choices, your phrasing, your stack of the day;
• “The agent” is co-constituted by the prompt, the routing, the reads, the refusals, the temperature of the moment;
• “Project X” is not an essence: it is actualized by what the apparatus can read, link, forget or hallucinate right now.

That is why two “identical” sessions on paper produce different worlds. This is not merely a reliability bug. It is the ontological structure of a system without a complete apparatus.

III. Consequences for designing agentic systems

If intra-action is taken seriously, the market’s fundamental question (“which is the best model?”) becomes secondary to:

Which apparatus organizes the emergence?

For Barad, an apparatus is not a machine in the hardware sense alone. It is the whole material-discursive set: prompts, permissions, interfaces, metrics, dossiers, handoff rules, memory windows, evaluation criteria. Each element takes part in what can appear as agent, task, context, failure, success.

1. The bare chatbot is an incomplete apparatus

A chatbot with no persistent dossier, no explicit permissions, no traces, no human-resume rituals is not “weak.” It is ontologically minimal: it produces poor intra-actions, where only the turn of speech counts, and where context dies when the tab closes.

2. Memory is not a “plus”

Memory (files, logs, distilled memories, handoffs) does not enrich a pre-existing entity. It stabilizes cuts so that certain intra-actions can repeat with enough continuity to deserve the name project or organization.

3. Agency is distributed — without abolishing human response-ability

Barad displaces agency: it is not the possession of a central human subject. It is enacted in the field. This does not mean “the human is no longer responsible.” It means: responsibility (response-ability, the capacity to respond to what emerges) concerns the design and revision of the cuts — not the fiction of total control after the fact.

This is the ethical core of Agentic Realism: refusing the gadget illusion (“yet another app”) and the magic illusion (“the AI takes care of everything”) in order to build revisable systems, where the human keeps taste, care and direction, and where the agents absorb the repetitive noise.

IV. From concept to practice

I don’t claim every SME must read Barad before deploying an agent. I claim that the bad failures of the market — lost context, opaque automations, three SaaS tools contradicting each other — are often failures of unthought intra-action: fictively stable entities were connected without designing the apparatus.

The practice I develop in Lausanne under Agentic Realism starts by mapping where the thread breaks (a context diagnostic), then by sprints that materialize assumed cuts: what the agent reads, what it never touches, how the human takes back over, what is written to survive a model change.

This is not “applied philosophy” in the decorative sense. It is an audit criterion:

• If you can’t describe how “the project” is actualized in the session, you don’t have a system — you have a conversation.
• If you can’t revisit your boundaries (permissions, folders, routing), you naturalize political cuts as if they were “just technical.”

V. Stop reflecting, start diffracting

Intra-action is the doorway. It opens onto other concepts from the same corpus — agential cuts, diffraction, response-ability — developed across the Agentic Realism series.

If this shift feels abstract, ask yourself: the last time an AI tool “answered well,” what was missing the next morning? The answer is almost always an apparatus breakdown, not a “bad answer.”

VI. From apparatus to living architecture: DNA Dossier, ExoBrain, Torus

Intra-action is not only a philosophical concept. It imposes a concrete architecture. If entities emerge in the apparatus, then the apparatus must be designed to preserve what matters: the DNA of a project, a person, an organization.

The DNA Dossier is the base layer: the history, the decisions, the values, the particular vocabulary, the constraints, the fragile intuitions, the relationships. It is not a dead file. It is an active memory that is re-read, updated, and re-linked.

When this DNA Dossier becomes structured, interconnected and operable, it becomes an ExoBrain: an external second brain inspired by the LM Wiki pattern (Karpathy) — persistent Markdown pages, entities, concepts, sources, cross-links, contradictions and evolving syntheses. Knowledge does not vanish between two sessions. It composes. It improves over time.

The Torus is the interface and architecture that circulates this context: it links the DNA Dossier, the files, the tasks, the automations, the agents, the permissions, the logs, the human decisions. A torus: what leaves comes back transformed, what circulates feeds the whole. It makes visible what the system knows, does, blocks, and where the human must decide.

This is syntropy in action: a gradual convergence between intent, living memory and concrete action. The vital impetus (Bergson) stays preserved because the apparatus is not rigid: enough structure for continuity, enough openness for creative variation.

The agents come after. They are organs of action within this living context, not the center. An agent without a DNA Dossier produces fast but loses the thread. An agent inside an ExoBrain can act rightly.

VII. Conclusion: keep the thread

To hold over time, intra-action demands visible agential cuts, diffraction (reading the differences rather than reflecting a single answer), and response-ability (the human capacity to respond to what emerges, with traces and resumes).

You do not use AI.
You enter an apparatus that makes emerge, each time, a partial-you, a partial-agent, a partial-project — and that keeps context alive through the DNA Dossier, the ExoBrain and the Torus.

The quality of an agentic system is measured by the richness and the endurance of these intra-actions over time — not by the polish of the last turn’s answer. That is the true durable advantage in a world where everyone will have the same models: context preserved, linked, maintained, alive enough to evolve without dissolving.

Sources and references

Primary philosophical source

Karen Barad — Meeting the Universe Halfway: Quantum Physics and the Entanglement of Matter and Meaning (Duke University Press, 2007). Concepts: agential realism, intra-action, agential cuts, diffraction, apparatus. The application to AI is an interpretive extension by the Agentic Realism project, not a thesis of the book.

Shorter entries: Barad’s articles and interviews; syntheses in STS (Science and Technology Studies).

Agentic Realism development

• Agentic Realism — Manifesto (2026): DNA Dossier, ExoBrain, Torus, LM Wiki, intra-action, agential cuts, diffraction, syntropy, response-ability, living context.
• Agentic Realism — Core Doctrine (2026): intra-action, cuts, diffraction, entropy/syntropy, human layer, vital impetus.
• Agentic Realism README (2025–2026): the brand-name ↔ Baradian agential realism link.
• Baradian-framework synthesis (Barad + Bergson + systems vocabulary).

Bergson (continued)

Henri Bergson — Creative Evolution (1907), the vital impetus (élan vital).

Syntropy / entropy

An interpretive vocabulary within Agentic Realism (context debt, logs, convergence of intent ↔ memory ↔ action). Syntropy is not a Baradian term.