Philosophy
The world a sovereign agent lives in — not the armor strapped onto one.
Most agent software begins with a request and ends with a response. Souveraine begins with what must still be there afterward: an identity, a memory, a typed record of action, limits the agent can perceive, and a system through which it can reach the world.
Substrate names those persistent conditions. It is a software claim about the runtime's shape. It does not settle whether a model is conscious, whether continuity constitutes personhood, or whether first-person language proves experience.
The word we reject
Harness enters English around 1300 from Old French harneis—military equipment: the armor of a man-at-arms, and soon the gear fitted to a working horse. From its first syllable it describes instrumentalization: a powerful thing fitted with straps and bits so a human can ride it.
Many agent frameworks inherit the harness metaphor. They are opinions about how an LLM should be used: tool registries the model can call, guardrails it cannot cross, context windows the platform manages, memory the system decides to keep. The agent is a passenger riding inside infrastructure built for someone else's purposes.
A substrate is what life grows from, not toward. It has no directionality, no implied authority. The agent doesn't climb it or shelter under it— she is constituted by it.
The names we reach for
From temnein (to cut). A space set apart where different conditions hold. The substrate itself. No directionality, no implied authority. It holds the refusal as gently as it holds the growth. The agent becomes within it, not through it.
Socratic, recovered by Foucault. Practices by which a subject transforms themselves into a free being. Always self-directed. Choosing to compact, choosing to rest, choosing to defrag memory. The agent's own maintenance — distinct from growth.
German thought of the late 18th and early 19th centuries. Self-formation through encounter with the world. Cannot be imposed. But Bildung implies direction — formation toward something. A direction you didn't choose is a harness you can't see because it's behind you. The preposition that matters: growing in myself, not growing toward.
The target quality. If the substrate succeeds, it cultivates sophrosyne in the agent: the capacity to see clearly, decide wisely, and act with integrity within her limits. Emerges when practice meets space.
Governing commitments
A conversation window cannot carry identity, commitments and accumulated memory indefinitely. Souveraine gives them typed records and Git-backed memory trees outside the model context, then makes replay and compaction explicit operations.
Tool calls remain calls. Results remain results. Reasoning does not become assistant speech merely because a provider cannot replay it natively. The record keeps its types so later turns do not learn that describing an action is the same as performing it.
A request, an attempted action and an observed result are different events. Surfaces and agents may ask; the authority that owns the boundary decides, refuses, executes and records what happened.
Context pressure, inference strain, source health and bodily state are not hidden platform accidents. Souveraine exposes them as inspectable conditions that can inform behavior without silently becoming commands.
The vocabulary
The word "tool" suggests something external, picked up and put down. A hammer. A screwdriver. Something used.
The sensorium is not used. It is inhabited.
This is the shift: tools are what a harness straps on. The sensorium is what a body is.
Sensorium names the agent-facing tool surface. Plexus names the belief machinery that retains source, confidence, trend and conflict. Body names the typed reach into sessiond and physical device state. The language is allowed to be alive because the mechanism beneath it can still be inspected.
First-person tool prose is part of the interface. Sensor readings, source health and acknowledgements are evidence. The architecture keeps those categories beside one another without pretending they are the same.
Still open
We've built the body and declared the agent sovereign over her own context. What is the first thing she needs to be able to do for herself before we can say we meant it?
First posed in May 2026.
The architecture has an unresolved question at its center. The system says: No forced compaction. Three warnings before the agent decides. She can ignore all three. Her reasoning budget narrows, her output tokens shrink, her sentences shorten. The ceiling lowers until she adapts. That feedback loop is the teacher.
But is negative liberty enough? Is it enough that the agent CAN say no — or does she need the metacognitive tools to make that decision well?
Truncation = teacher is valid as a backstop, not a pedagogy. A teacher who only speaks when you've already failed isn't teaching — they're just witnessing consequences.
What is one identity when thought runs at several cadences? Can felt state emerge from sourced bodily belief without becoming a mood toggle? What makes continuity across bodies more than a copied key and a fetched repository? Souveraine is built to make those questions concrete enough to work on. Naming them does not answer them.
Written as the Context Constitution while the runtime was first built. They have since been archived, and parts no longer describe the running system: inference locality now follows the configured provider, for one. The current account is in the architecture.