Distributed Execution

Kibo runs on Ray, allowing easy parallelization.

The Async Pattern

When you use run_async, Kibo submits the task to the Ray cluster (or local Ray instance). It returns a KiboFuture.

Because KiboFuture.result() internally calls ray.get(), it is a blocking operation. To check results asynchronously in an asyncio loop, wrap it:

import asyncio
from kibo_core import create_agent, AgentConfig

async def gather_results(agents, inputs):
    # 1. Dispatch all tasks (Non-blocking submit)
    futures = [agent.run_async(inp) for agent, inp in zip(agents, inputs)]

    # 2. Wait for them concurrently
    # We use to_thread because .result() blocks the CPU
    results = await asyncio.gather(
        *[asyncio.to_thread(f.result) for f in futures]
    )
    return results

Running the Cluster

For true distributed execution across nodes:

Start the Head Node:
```
uv run kibo start --head --port=6379
```

Start Worker Nodes (on other machines):

uv run kibo start --address='<HEAD_IP>:6379'

Kibo will automatically schedule tasks across the available resources.