kcastle-agent

A minimal OpenAI Responses Agent core with optional harness infrastructure.

from openai import AsyncOpenAI

from pathlib import Path

from kagent import Agent, Session

session = Session.create(Path.home() / ".kcastle" / "sessions")

agent = Agent(
    client=AsyncOpenAI(),
    model="gpt-5.6-sol",
    instructions="You are a helpful agent.",
    state=session.state,
    commit=session.commit,
)

async for event in agent.run("Hello"):
    print(event)

Agent is the brain: it owns causal State, the loop, steering, follow-ups, cancellation, events, and compaction. kagent.harness surrounds it with Session persistence, an explicit Env, and executable ToolRuntime adapters. User interfaces render events and provide approvals only. When one model response contains multiple function calls, execute_tool may run concurrently; custom executors must be safe for concurrent use.