The problem
Every agent interface reinvents how to stream a run. Ad-hoc event formats make runs hard to render, record, replay and debug, and they lock the interface to one framework.
Why it matters
People supervise agents through what they can see. A run that can be followed, paused and replayed is a run that can be trusted with more.
Research question
What is the smallest event vocabulary that lets any agent stream a legible run to any client?
Approach
- A JSON-schema protocol over server-sent events with a small set of typed events: token, tool use, thinking, progress and done.
- A Python emitter and a React hook as reference implementations.
- Recording and replay of runs from saved event logs.
Methodology
- The schema and both reference implementations are tested in the repository.
Findings so far
Used in production by agentFast, which builds on the protocol for live streaming and adds events that a standard client can safely ignore. No formal evaluation has been run.
Limitations
- Early version, with no compatibility guarantees yet.
- Adoption so far is within Einstein Labs products only.
Artifacts
- Repository · MIT