Agent frameworks keep appearing. LangGraph talks about Checkpoint; OpenAI talks about Thread and Run; A2A talks about Task; AG-UI talks about Event; Deep Agents adds Todo, Subagent, and Virtual Filesystem.
The names change, but the underlying problems do not. Every serious Agent runtime must answer the same questions:
- How is a task started?
- What context does it carry?
- How can progress be observed?
- How can it pause, resume, retry, or be cancelled?
- Where do final artifacts live?
- How can state be recovered after interruption?
This article argues that the stable layer is not a specific framework API. The stable layer is a set of protocol objects and lifecycle operations that repeatedly appear across frameworks.
The Six Objects
A production-grade Agent protocol needs at least six conceptual objects.
| Object | Plain meaning | Question it answers |
|---|---|---|
Thread / Session | Long-lived context | Which user or task context is this? |
Run / Task | One concrete execution | What exactly is being executed now? |
Step | Observable unit of work | Which model, tool, or sub-Agent step happened? |
Event | Runtime progress signal | What changed during execution? |
Artifact | Durable output | Where is the result and which run produced it? |
Checkpoint | Recoverable state snapshot | Where can execution resume after failure? |
Around these objects, the protocol also needs operations such as stream, interrupt, resume, cancel, and retry.
Protocol and Runtime
Protocol is the external boundary of a runtime. Runtime is the internal implementation that fulfills the protocol.
This distinction matters because many framework debates confuse three layers:
- concrete standards such as A2A, AG-UI, LangChain Agent Protocol, AITP, or ACP;
- general protocol objects such as Thread, Run, Step, Event, Artifact, and Checkpoint;
- runtime capabilities such as persistence, interrupt recovery, tracing, permission control, and evaluation.
The article focuses on the second layer. Specific standards and frameworks are evidence, not the main point.
Why This Matters
The core of an Agent runtime is not a model call. It is task lifecycle management.
A toy Agent can call a model and tools. A production Agent must define task boundaries, manage state, expose progress, support interruption, recover from failures, and produce auditable artifacts.
That is why long-term attention should go to protocol boundaries and runtime abstractions rather than API names. Frameworks will change. The lifecycle problems will not.
Reading Guide
Use this draft as an English entry point to the Chinese source article. The source contains the full framework comparison, diagrams, and detailed runtime analysis.