Source: arxiv.org/abs/2610.21538
Under the protocol, agents of different kinds send small JSON events over HTTP POST to a status server. The server then exposes the aggregated state to observers such as dashboards over Server-Sent Events, so changes reach them as they happen.
Because a crashed, hung or compromised agent cannot be relied on to report its own failure, the server also infers liveness from missed heartbeats and can flag agents that are running but no longer making progress. It accepts observations from supervisors, guardrails, operators and peers, and some of these take precedence over what an agent says about itself while others are shown alongside it.
State is tracked along separate dimensions: lifecycle, health, task, floor and governance. An agent can therefore be online and degraded while working on a task and waiting for the floor at the same time. Version 0.1 borrows task states from the Agent2Agent protocol, floor states from the Open Floor Protocol and attribute conventions from OpenTelemetry GenAI semantic conventions.
Why it matters if you build agents
Developers running multi-agent systems could get a shared vocabulary for health and progress signals instead of defining their own status fields. Heartbeat-based liveness combined with outside observations can catch agents that stall silently, which self-reports alone would miss.
Because this is a working draft, its states and field names may change, so anyone building against it should treat the implementation as provisional.
Get stories like this every weekday
One short email with the day's agent tooling news.