Original skill document
Rendered from the source repository. Headings, examples, code, tables, links, and referenced images are preserved.
Deep Agents Python quickstart
Follow the live docs — do not invent an alternate API from memory:
https://docs.langchain.com/oss/python/deepagents/quickstart
Fetch that page (Docs MCP or HTTP) and implement the research-agent shape it shows (create_deep_agent, research system prompt, invoke with a research question like “What is LangGraph?”).
Local setup constraints
Apply these on top of the quickstart (they keep setup minimal and model-agnostic):
- Ask which provider/model to use. Showcase that Deep Agents are model-agnostic. Suggested prompt:
Which model should this agent use? Pass aprovider:modelstring — e.g.openai:gpt-5.5,anthropic:claude-sonnet-5,google_genai:gemini-3.5-flash. Default if you're unsure: `anthropic:claude-sonnet-5`. We'll use that provider's built-in web search (no separate search API key).
- Create a new directory (e.g.
deep-agent/) and do all work there — do not pollute the open project.
- Do not use Tavily (or any second search vendor). Replace the quickstart's
internet_search/ Tavily tool with the chosen provider's built-in web search. Look up the current tool shape on that provider's LangChain chat docs (examples as of writing — re-check if needed):
| Provider | Built-in search tool |
|---|---|
| Anthropic | {"type": "web_search_20260209", "name": "web_search", "max_uses": 5} |
| OpenAI | {"type": "web_search"} |
{"google_search": {}} |
Prefer Anthropic / OpenAI / Google so provider search is available. Only secret: that provider's API key in .env (gitignored). Skip LangSmith tracing unless they ask.
- Install
deepagents(+python-dotenv) and the provider package for their model — nottavily-python.
- Run the research example, show output, then stop. Point to
deep-agents-core/ customization / Managed Deep Agents for next steps.
