Getting Started with OpenAI Agent Kit¶

OpenAI Agent Kit packages opinionated utilities for building production-ready AI agents—wrapping model calls, tool routing, and evaluation in a way that scales from quick prototypes to real apps.

This guide walks you through installing Agent Kit, scaffolding a minimal agent, wiring tools, and running automated checks so you can ship confidently.

Prerequisites¶

Python 3.10+
An OpenAI API key stored in OPENAI_API_KEY
uv or pip for dependency management

1. Install the Agent Kit¶

Create and activate a virtual environment, then install the core package and CLI helpers:

uv venv
source .venv/bin/activate
uv pip install openai-agent-kit openai

Prefer pip? Swap the last line for pip install openai-agent-kit openai.

2. Scaffold an Agent project¶

Generate a starter layout with the built-in CLI:

agentkit init my-agent
cd my-agent

The CLI creates a modular project:

my-agent/
  agentkit.toml        # project metadata and model defaults
  main.py              # runnable agent entrypoint
  tools/               # register custom tools here
  tests/               # agent regression tests

3. Configure the agent¶

Open main.py and register the core planner, tool router, and a short-term memory store:

from agentkit import Agent, OpenAIChatModel
from agentkit.memory import WindowMemory
from agentkit.tools import ToolRegistry

from tools.github import create_issue

model = OpenAIChatModel(model="gpt-4.1")
tools = ToolRegistry([create_issue])
memory = WindowMemory(limit=6)

agent = Agent(
    model=model,
    tools=tools,
    memory=memory,
    system_prompt="You triage GitHub issues and draft helpful replies.",
)

if __name__ == "__main__":
    agent.run()

This setup gives the agent a task-specific system prompt, bounded conversational memory, and a single GitHub issue tool.

4. Add a custom tool¶

Create tools/github.py to wrap business logic or API calls:

from agentkit.tools import tool
from github import Github

@tool(name="create_github_issue", description="Create an issue in the lidenlab/site repo")
def create_issue(title: str, body: str) -> str:
    client = Github()
    repo = client.get_repo("lidenlab/site")
    issue = repo.create_issue(title=title, body=body)
    return f"Created issue #{issue.number}: {issue.title}"

Decorating the function makes it discoverable by the planner and automatically converts docstrings into tool descriptions.

5. Run and iterate¶

Start the agent locally:

python main.py

The CLI streams reasoning steps, tool invocations, and final answers. Add --debug to log raw JSON payloads when troubleshooting.

6. Test the agent contract¶

Agent Kit ships with pytest fixtures for replay testing. Add a regression test to tests/test_issue_triage.py:

from agentkit.testing import AgentHarness

harness = AgentHarness("fixtures/issue_triage.yaml")


def test_triage_flow():
    result = harness.run("Summarize open bug reports")
    assert "summary" in result.data

Replay tests keep tools stable and prevent regressions as you tweak prompts or upgrade models.

Next steps¶

Explore the Agent Kit docs for deployment recipes.
Swap WindowMemory for vector memory to recall longer histories.
Add observability with agentkit.telemetry to trace token usage and tool latency.

With this foundation, you can iterate from a single scripted tool to a full workflow orchestrator backed by Agent Kit.