Agential cuts: every digital boundary is a political decision

I. Prologue: the invisible boundaries

An agentic system looks transparent when it answers politely. Beneath the surface, it performs cuts: it includes certain files, excludes others; allows certain tools, forbids the rest; treats one email thread as “context” and another as “off topic.”

These boundaries are not technical details. They are decisions about what exists for the agent — and what does not. The physicist and philosopher Karen Barad calls them agential cuts.

II. Agential cuts (Karen Barad)

In Meeting the Universe Halfway (2007), Barad shows that any measuring device — any apparatus — does not merely observe: it cuts reality. What is made visible becomes measurable, actionable, debatable. What is excluded from the apparatus simply does not enter the world produced by that configuration.

Transposed to digital systems (an interpretive extension, not Barad’s thesis): every architecture choice is an agential cut. Listing three folders in a skill says: “the project lives here.” Forgetting the CRM says: “the client does not exist for this session.”

Cuts are not intrinsically bad. Without them, no targeted action. The problem begins when they are implicit, non-revisable, or presented as “neutral.”

III. Concrete examples in AI stacks

• RAG scope: which documents are indexed? Internal notes, or only the public site?
• Allowed tools: can the agent send an email, modify a file, call a payment API?
• Memory: what is distilled into long-term memory (files, structured notes) vs what stays in ephemeral transcripts?
• Handoff: what does one agent “see” when another hands over to it?
• Human roles: who validates, who signs, who can cancel an autonomous action?

Two organizations with the “same” model and the “same” prompt produce different worlds because their cuts differ. This is a direct consequence of intra-action: the apparatus co-constitutes the agent and the task.

IV. Naturalizing the cut (“it’s just an assistant”)

The surrounding discourse often hides the cuts: “the AI reads your data,” “the agent acts for you.” In practice, it reads what you plugged in, and acts where you opened a door.

To naturalize is to pass off a political configuration as a technical given. The result: when the system fails, we blame the model or the user — rarely the badly drawn boundary.

V. Agentic Realism: legible, revisable, assumed cuts

Agentic Realism, Response-able AI does not claim to remove cuts — that is impossible. It aims to make them explicit: documented (skills, permissions, DNA Dossier), auditable (traces, logs), modifiable when the business changes. The Torus makes these boundaries visible and steerable in the context-circulation interface.

An assumed cut can be debated as a team: “Why doesn’t the agent have access to the contracts?” A hidden cut only produces frustration and context hallucinations.

Response-ability (the capacity to respond) starts there: knowing where the system drew the line, so you can come back with human judgment — taste, care, direction.

VI. Quick audit (five questions)

1. Which file paths can the agent read — and which are absent by design?
2. Which external actions are possible without human validation?
3. Where is long-term memory stored — and who can modify it?
4. What is excluded from “context” by default (noise, PII, old versions)?
5. Who can reconfigure these boundaries without redeploying the whole stack?

If you can’t answer in one page, your cuts don’t yet serve the human — they serve themselves, unnoticed.

VII. From the cut to the living system: DNA Dossier, ExoBrain, Torus

Every agential cut draws the perimeter of the DNA Dossier: what enters the project’s active memory, what stays excluded. A badly thought-out cut creates entropy (lost context, fragmentation). An assumed cut enables syntropy: a gradual convergence between living memory, intent and concrete action.

The DNA Dossier becomes an ExoBrain when the cuts are structured into interlinked pages (LM Wiki style, Karpathy): entities, concepts, sources, contradictions, evolving syntheses. Knowledge composes across sessions instead of starting from zero.

The Torus links the DNA Dossier, files, tasks, automations, permissions, logs and human decisions. It shows what the system knows, does, blocks — and where the human must decide. Agents are organs of action there, not the center of the apparatus.

VIII. Conclusion

Every digital boundary is a decision about possible worlds. Making them visible is the precondition for any responsible AI that holds up in an SME, a practice, a busy life. It is the foundation of the Agentic Realism architecture: living context preserved, explicit cuts, the human keeping taste and direction.

Next in the series: diffraction — reading the differences and interference between sources and angles rather than a single mirror answer. Then entropy and syntropy, response-ability in depth, and living context as the durable advantage when everyone has the same models.

Sources and references

Karen Barad — Meeting the Universe Halfway (Duke University Press, 2007): agential cuts, apparatus, agential realism. Application to AI systems: an interpretive extension by Agentic Realism.

• Agentic Realism — Manifesto (2026): DNA Dossier, ExoBrain, Torus, LM Wiki, intra-action, agential cuts, diffraction, syntropy, response-ability, living context, vital impetus.
• Agentic Realism — Core Doctrine (2026): cuts, diffraction, human layer, entropy/syntropy.
• Article Intra-action (2026): co-constitution of apparatus / agent / project.
• Baradian-framework synthesis (Barad + Bergson + systems vocabulary).