Status: exploration. A design document, no implementation, no timeline. It is published early so the shape of the idea is on the record and can be argued with.
An agent session today is one opaque, growing state machine. Reasoning, memory, context, tool execution, scheduling, governance, identity and permissions all live inside the same conversation, and every one of them gets worse as it grows. The result is familiar: context that only expands, state you cannot inspect, runs you cannot reproduce, and a hard dependency on whichever model was used.
Operating systems solved the equivalent problem for ordinary programs by defining primitives — processes, files, permissions, signals, sockets — and letting everything else be built on top. Agent systems have no such layer. They have frameworks.
This is an attempt to describe the missing one:
-
Sessions are disposable. They start, do work, produce artifacts, end.
-
Models are workers. They execute reasoning. They own no state, no identity, no governance, and they are replaceable.
-
Context is virtual memory, not a transcript — pages, handles, overlays, materialised on demand from a much larger semantic space.
-
Artifacts are the system of record, immutable, versioned, replayable.
-
Governance and capabilities are runtime services, not paragraphs in a prompt asking an agent to behave.
-
spec/01-philosophy.md — Volume I: the mentality. Why every harness rewrites the same runtime, why inventing a primitive is a bad sign, why no conversation is stored anywhere, and why Linux is underneath rather than in the way.
-
spec/00-vision.md — motivation, invariants, architectural direction, and the eighteen volumes this is meant to grow into.
-
spec/02-architecture.md — Volume II, first sketch: the layer model, what is in scope (L2–L4) and what is deliberately not, and where models sit (L2, as drivers).
-
spec/03-kernel.md — Volume III, first sketch: an argument that six primitives are enough, and that the scheduler, interrupts, events and drivers are not among them.
The invariants above are not whiteboard ideas. Most of them are already running in a private multi-agent system that has been in daily use for months: governance gates that block a hand-off until evidence exists, work represented as artifacts rather than chat history, role-scoped capabilities enforced at the tool layer, checkpoints and replayable outcomes, and a measurement habit that has repeatedly proven the intuitions wrong.
That system is not this specification, and it is not public. But it is the reason this document describes an architecture rather than proposes one: the expensive lessons have already been paid for somewhere.
Existing systems are described here, not evaluated. A design read as a mistake teaches nothing; the same design read as a rational answer to a constraint tells you what the constraint was — and that is the part worth having. The monolithic agent session is not a failure, it is what you build when there is no runtime to build on.
Not a framework. Not an orchestration layer. Not a competitor to MCP, to any model vendor, or to Linux. Where an existing standard fits a primitive, the specification should adopt it rather than restate it.
One volume at a time, and only when there is something to say. Empty chapter stubs are how specifications die, so a volume becomes a file when it has content. Volume order will follow whatever gets implemented or tested first, not the numbering.
Discussion belongs in issues. Disagreement with an invariant is the most useful thing anyone can contribute at this stage — they are load-bearing, and they are cheaper to break now than in Volume XV.
The specification text is CC BY 4.0. Any code in this repository is MIT.