Build a browser agent with Microsoft Agent Framework

examples/microsoft-agent-framework

Contributors: Jun RyuUpdated May 11, 2026

Microsoft Agent Framework is Microsoft's 1.0 agent runtime (Python + .NET), the successor to AutoGen and Semantic Kernel. Provider-agnostic chat clients, function tools as plain Python callables, and built-in MCP / A2A interop.

This starter wires a Steel browser into the framework's @tool decorator pattern and points the agent at GitHub Trending.

from agent_framework import Agent, tool
from agent_framework.openai import OpenAIChatClient

agent = Agent(
    client=OpenAIChatClient(model="gpt-5-mini"),
    name="SteelBrowserAgent",
    instructions="You operate a Steel cloud browser via tools. ...",
    tools=build_tools(page),  # navigate, snapshot, extract
)

result = await agent.run(
    "Go to https://github.com/trending/python ..."
)
print(result.text)

agent.run runs the model loop until the agent stops calling tools. result.text aggregates every text content item across the run's messages; result.messages is the full transcript including function_call and function_result contents if you want to inspect what happened.

Tools are plain Python functions decorated with @tool. The framework infers the JSON schema from the type hints (Annotated[str, Field(description=...)]) — or you can pass a Pydantic model via schema=:

@tool(
    name="navigate",
    description="Navigate the open session to a URL and wait for the page to load.",
    approval_mode="never_require",
)
async def navigate(
    url: Annotated[str, Field(description="Absolute URL to navigate to.")],
) -> dict:
    await page.goto(url, wait_until="domcontentloaded", timeout=45_000)
    return {"url": page.url, "title": await page.title()}

Tools are built inside build_tools(page) so each run closes over its own Page — no module globals, safe to run concurrently.

Run it

cd examples/microsoft-agent-framework
cp .env.example .env          # set STEEL_API_KEY and OPENAI_API_KEY
uv run playwright install chromium
uv run main.py

Get keys from app.steel.dev and platform.openai.com. Each tool call prints its latency so you can see where time is going.

Your output varies. Structure looks like this:

Steel + Microsoft Agent Framework Starter
============================================================
Session: https://app.steel.dev/sessions/ab12cd34...
    navigate: 1612ms
    snapshot: 487ms (3821 chars, 48 links)
    extract: 394ms (3 rows)

Agent finished.

1. owner/repo — https://github.com/owner/repo (1,240 stars)
   ...

Releasing Steel session...
Session released. Replay: https://app.steel.dev/sessions/ab12cd34...

A run takes ~20 to 40 seconds and 5 to 10 agent turns on GitHub Trending. Cost is a few cents of Steel session time plus OpenAI tokens. The finally block in main closes Playwright and calls steel.sessions.release() so Steel stops billing per-minute.

Make it yours

• Swap the model. Change the OpenAIChatClient(model=...) argument. 'gpt-5', 'gpt-4o', and the chat-completion variants all work. For Anthropic or Azure OpenAI, swap the client: from agent_framework.anthropic import AnthropicChatClient or from agent_framework.azure import AzureOpenAIChatClient.
• Swap the task. Change PROMPT and the INSTRUCTIONS system text. Tools stay the same; the agent re-plans against the new shape.
• Structured output. Pass a Pydantic model via response_format in default_options (or the per-run options). The agent's final message will be validated against the schema.
• Stream the answer. Use async for update in agent.run(prompt, stream=True): to stream tokens as they arrive. Tool calls happen behind the scenes; update.text carries the partial answer.
• Run agents in parallel. Construct a session+page+agent per task and asyncio.gather(*[agent.run(...) for agent in agents]). Each closure captures its own page; nothing is shared by accident.
• Approval gates. Drop approval_mode="never_require" on a sensitive tool and the framework will pause for an approval response before invoking it — useful for destructive actions (form submits, purchases).
• Multi-agent workflows. The framework ships a Workflow / GroupChat API for graph-based multi-agent orchestration; one Steel-driven Agent becomes a node in a larger graph.

Related

Steel + Claude Agent SDK (Python) · Steel + Pydantic AI · Microsoft Agent Framework docs

Related recipes