restate-agent

Restate-backed, Pydantic-typed agent runner inspired by PydanticAI's TemporalAgent.

Quickstart

from pydantic import BaseModel
from restate_agent import RestateAgent

class Answer(BaseModel):
    text: str

class DummyAgent:
    async def run(self, prompt: str, deps=None, tools=None) -> Answer:
        return Answer(text=f"Echo: {prompt}")

agent = RestateAgent(DummyAgent(), output_type=Answer)
result = await agent.run("hello")
print(result.text)

Agent contract

RestateAgent expects a runner with this shape:

async def run(self, prompt: str, *, deps=None, tools=None) -> OutputT: ...

The agent is responsible for using the registered tools (if any) and returning the output type you pass to RestateAgent.

Durable execution

Durable execution uses an activity executor (e.g., Restate workflow context) to run the non-deterministic agent call in an activity while keeping orchestration deterministic.

from restate_agent import RestateAgent
from restate_agent.runtime import RestateActivityExecutor, create_durable_runtime

# ctx is your Restate workflow context
runtime, handler = create_durable_runtime(
    agent,
    lambda _handler: RestateActivityExecutor(ctx),
)
# handler is the activity entrypoint you register with Restate
result = await agent.run("hello", runtime=runtime)

Optional integrations

• restate SDK: pip install "restate-agent[restate]"

See docs/ and examples/ for more detailed usage.

Build & test

python -m pip install -e .[restate]
python -m pytest
python -m build
basedpyright

Notes

TemporalAgent compatibility: this library mirrors PydanticAI's TemporalAgent ergonomics but maps workflows/activities to Restate and uses create_durable_runtime(...) to wire the activity boundary.

Durable execution expects JSON-serializable deps, tool payloads, and outputs.