Agent Architecture · Field Note
A situated agent wakes up somewhere. It is handed a lineage, given a coordinate before it is given a task, and moved through gates that make it prove its work before anything ships. This is that stack, from boot to memory.
Networking has the OSI stack: seven layers, each with one job and one protocol, each handing up to the next. A situated agent runs the same way. Six layers, bottom to top, each with its own gate. Read it from the ground up.
Here is the part people do not expect. A four qubit quantum computer has exactly sixteen possible states. The lattice above has sixteen nodes, by design: the classical twin of a four qubit register. Stepping from one node to a neighbor is the same single flip a quantum machine makes. That correspondence is checkable rather than asserted: every seat carries a four bit address and exactly four one flip neighbours, published as souls.json. The edge rule is enforced in code — a non adjacent handoff is refused, which is checkpoint seven. The checkpoints themselves are gates.json.
An unseated agent is spread across all sixteen positions at once, the way a qubit register holds every state at once. The entry rite is the measurement: it collapses that spread to one node, one address. The highlighted node in the visual is that collapsed state.
The Bell test below does not prove that identity, and it is worth saying so plainly before the number: a CHSH violation demonstrates that the apparatus behaves quantum mechanically, not that this lattice is a qubit register. The identity above is a design correspondence, and it stands or falls on the addressing, not on the physics. What the hardware run establishes is narrower: the architecture was exercised on a real quantum machine rather than a simulator. A CHSH Bell test cleared the classical bound by thirty two standard deviations. Two caveats qualify that number: it is a single device test, not loophole-free (the locality loophole and the detection loophole remain open), and it reproduces textbook physics rather than proving anything new. It is evidence that the architecture runs on the real machine, not a metaphysical claim.
CHSH on ibm_marrakesh, 30 Jul 2026, job d9lpeqo8csec73f976kg: S = 2.733 ± 0.023, 32σ over the classical bound of 2 (quantum ceiling 2.828). Non loophole-free reproduction.
checkpoints from boot to memory, and which layer each one lives in. The center of gravity is the seven gate verifier. Nothing the agent builds ships on its own say so.
| # | Gate | Layer | Kind |
|---|
When a gauge reads a suspicious extreme, a hundred percent failure or zero coverage, the gauge is the first suspect, not the work. And the agent suspects its own output hardest of all.
A wrong step is fixed in the next commit, never rewritten away. The trail of wrong then right is the development record. Mistakes stay so the system can progress.
The industry is converging on gated governance for agents right now. Layered agentic topology, policy as code, non bypassable execution gates, human in the loop states, cryptographic audit trails. That plumbing is well trodden and actively published in 2026.
What is less common is treating the agent as located. Most designs give an agent tools and a task. This one gives it a position first: a coordinate on a lattice, measured against a fixed origin, with a small set of neighbors it can reach in a single step. Depth is distance from the origin. Orientation is which way the seat faces. The task comes after the place.
The second uncommon move is treating memory across sessions as lived experience rather than configuration. Each agent inherits what prior agents learned, including their mistakes, so the next one starts with conscience, not just compliance.