xr-ai-runtime

xr-ai-agent-runtime provides typed, participant-scoped publication and fan-out for composable agents. It does not own models, planning, memory, tools, media, application state, or agent-created tasks. Refer to API Reference for the exact public APIs.

Agents and tools

An Agent owns private state and exposes ordinary finite Tool or streaming AsyncTool objects from xr-ai-tools. Direct callers execute those tools directly; model loops expose finite tools through a ToolSet. The runtime never wraps or redispatches tool calls.

from pydantic import BaseModel
from xr_ai_runtime import Agent, AgentRuntime
from xr_ai_tools import Tool

class Text(BaseModel):
    text: str

class TextAgent(Agent):
    def __init__(self) -> None:
        self.uppercase = Tool(
            "uppercase", "Uppercase text.", Text, Text, self._uppercase
        )
        super().__init__((self.uppercase,))

    async def _uppercase(self, request: Text) -> Text:
        return Text(text=request.text.upper())

runtime = AgentRuntime()
text = runtime.register("text", TextAgent())
async with runtime:
    result = await text.uppercase.execute(Text(text="hello"))

When one model sees tools from several agents, namespace only the model-visible catalog with ToolSet.namespaced(...). Participant identity remains a typed tool input supplied at the application boundary.

Publication and ownership

publish(topic, event) is asynchronous fan-out. It waits for all subscriber deliveries and then propagates failures. A subscriber that performs lengthy work must hand it to its own bounded queue and return promptly. The runtime owns delivery tasks only; an agent owns creation, synchronization, cancellation, and cleanup for its resources and background work.

Callbacks and tools may run concurrently. Agents protect shared mutable state with their own lock or queue so unrelated and streaming work is not globally serialized. Participant controls such as start, stop, and status are ordinary application tools; agent lifetime is not a model tool.

Topics use full Relay telemetry by default. High-cardinality transport topics may declare telemetry="none" when their consumer emits one semantic operation scope instead. Runtime publication and delivery scopes carry participant, message, correlation, parent-message, source, and subscriber metadata. A detached agent task whose lifetime exceeds its callback starts a fresh Relay scope stack and preserves logical correlation in metadata rather than retaining an ended delivery scope. Pass context=nemo_relay.fork_asyncio_context() to asyncio.create_task(), then enter nemo_relay.use_scope_stack(nemo_relay.create_scope_stack()) inside the detached coroutine before opening its operation scope or calling tools.