The hard part of a serious agent application is rarely the model call. It is everything around the conversation: threads that must outlive the runtime that served them, version pinning so a thread stays on the agent it started with, upgrades that land mid-conversation without dropping state, and an audit trail that can answer what the agent did, as whom, and when. Every team building agents ends up hand-rolling that machinery, and none of it is the product.
OASP — the Open Agent Session Protocol — standardises that layer. It rests on one idea: a conversation is not a session. The Conversation is the durable, user-facing thread. Sessions are the disposable, provider-side execution contexts it rides on. Hold those two apart and the operations every agent platform needs stop being in-house heroics and become protocol interactions: publish a Conversation onto a Deployment, migrate it to a new Session with lineage preserved, drain the old one cleanly.
A SMALL, CLOSED SURFACE
The whole standard is eight resources and five interactions. AgentDefinition, Deployment, Conversation, Session, Event, Principal, Credential, AuditEvent — nothing else. Identity and audit are not extensions: AuditEvent is normative, so “what did the agent do, as whom, when” is answerable by any conformant implementation. Conformance itself comes in three levels — client, server, adapter — with one kit to prove all three.
The hard part becomes a spec you adopt, not a system you maintain.
THE STANDARD CAME FIRST
OASP started as Loom. Looking at the popular LLM gateways, I kept hitting the same ceiling: every provider flattened into a chat-completions-shaped API, with the richness of a managed agent sanded off in the middle. I wanted the opposite — conversations as the core abstraction, each provider handled by its own implementation at the edges. And I worked the way I now always work: I described that shape to my AI pair, and it prototyped against the Anthropic API, wired into Loom, faster than I could have specified it on paper.
Partway in, we stopped and talked about what we were actually building. I pointed the AI at systems I had already built — LucidBrain, WorkSpec, Cortex and the rest of the Fieldstate portfolio had each solved pieces of durable conversations, identity, and audit in their own way — and we agreed the first best step was not another gateway. It was the contract those systems kept re-implementing. So the standard shipped first. Loom follows as its reference server, and the conformance kit exists so it will not be the only implementation: the ecosystem is meant to be more than one vendor.
WHAT THE PROCESS TAUGHT ME
This project is the clearest picture of how I build now. The prototype was nearly free — my AI pair turned description into working code as fast as I could describe it. What was not free was noticing we were building the wrong artifact first; that took the prior systems, the discussion, and a judgment call. AI has collapsed the cost of making things, which makes deciding what to make the whole job. The rest was subtraction: deciding what the protocol would refuse to carry so that what remains can be conformed to. Eight resources is not what I started with. It is what survived.