agentkits

agentkits is a lightweight async agent SDK for building tool-using agents on top of OpenAI-compatible Chat Completions, OpenAI Responses, and Claude.

The package keeps the runtime small and explicit: messages, models, tools, and agents are separate layers.

Core Ideas

Layer	Responsibility
message	Normalizes system, user, assistant, and tool messages across providers.
model	Wraps OpenAI-compatible Chat Completions, OpenAI Responses, and Claude behind one async interface.
tool	Registers Python functions as JSON-schema tools and dispatches tool calls.
agent	Provides ReAct, planning, handoff, agent-as-tool, and reflection-style runtimes.

ReActAgent is the core execution loop. PlanAgent plans first, then runs an internal ReAct executor. handoff(...) lets a router agent transfer a conversation to another agent through a tool marker.

Quick Start

import asyncio

from agentkits import OpenAIChatModel, ReActAgent, Toolkit, ToolResponse

toolkit = Toolkit()


@toolkit.tool()
def add(a: int, b: int) -> ToolResponse:
    """Add two integers.

    Args:
        a: First integer.
        b: Second integer.
    """
    return ToolResponse.from_value(str(a + b))


async def main() -> None:
    async with OpenAIChatModel.from_ali_env() as model:
        agent = ReActAgent(model=model, toolkit=toolkit)
        result = await agent.run("Use the tool to calculate 21 + 21.")
        print(result.text())


asyncio.run(main())

The larger demo in examples/08_reproducible_demo.py shows a full order-recovery workflow with a router agent, handoff, plan generation, tool calls, and final output.

Agents

Agent	Purpose
ReActAgent	Tool-calling loop.
ClassicReActAgent	Text-based Thought / Action / Observation loop.
PlanAgent	Plan first, then execute.
ReWOOAgent	Plan a tool DAG, execute observations, then solve.
ReflexionAgent	Act, evaluate, reflect, and retry.
SelfRefineAgent	Draft, review, and revise.

License

Apache-2.0