Building a research agent with LangGraph and MCP

1 Give each source a job

2 Ask what a tool can do

The Model Context Protocol (MCP) gives an application a common way to discover and call tools. A server describes the operations it offers and the inputs each needs. A client in our notebook asks for that list and passes it to the model. The model can then ask for a particular operation.

In the live notebook, the community NBER-CLI server runs on the reader’s computer. The official FRED connector runs at the St. Louis Fed. Both use the same protocol. The percentage calculation is a small Python function the agent can call directly. MCP architecture

3 Give the agent a way back

Each LangGraph node does one job. The decision node reads the original question and the history of calls and replies. If it requests a tool, the tool node runs it, appends the result and hands control back. If the agent has enough information, it writes its answer. This conditional edge is the branching arrow in the diagram.

def next_step(state: ResearchState) -> Literal["use_tool", "finish"]:
    if state["decision"]["tool"] == "finish":
        return "finish"
    return "use_tool"

4 Watch the tools change with the question

Choose a question. Each row is a call executed by the notebook. Open a row to inspect the arguments and the reply that informed the agent’s next decision.

search_papers {
  "query": "Generative AI at Work"
}

[
  {
    "id": "w31161",
    "title": "Generative AI at Work"
  }
]

get_paper_info {
  "paper_id": "w31161"
}

{
  "id": "w31161",
  "title": "Generative AI at Work",
  "version": "November 2023 working paper",
  "finding": "Following a staggered rollout of AI assistance, 5,179 customer-support agents resolved 14% more issues per hour on average when given access to the tool.",
  "url": "https://www.nber.org/papers/w31161"
}

search_series {
  "query": "U.S. nonfarm business output per hour quarterly index"
}

[
  {
    "id": "productivity",
    "title": "U.S. nonfarm business labor productivity"
  }
]

series_metadata {
  "series_id": "productivity"
}

{
  "id": "productivity",
  "title": "U.S. nonfarm business labor productivity",
  "measure": "Output per hour worked",
  "units": "Index, 2017 = 100",
  "frequency": "Quarterly",
  "seasonal_adjustment": "Seasonally adjusted",
  "vintage": "BLS third-quarter 2026 table",
  "url": "https://www.bls.gov/charts/productivity-and-costs/nonfarm-business-sector-indexes.htm"
}

series_observations {
  "series_id": "productivity",
  "start": "2022-10-01",
  "end": "2024-10-01"
}

{
  "observations": [
    {
      "date": "2022-10-01",
      "value": 110.219
    },
    {
      "date": "2024-10-01",
      "value": 116.978
    }
  ],
  "url": "https://www.bls.gov/charts/productivity-and-costs/nonfarm-business-sector-indexes.htm"
}

percent_change {
  "start_level": 110.219,
  "end_level": 116.978
}

{
  "percent_change": 6.132
}

search_papers {
  "query": "Generative AI at Work"
}

[
  {
    "id": "w31161",
    "title": "Generative AI at Work"
  }
]

get_paper_info {
  "paper_id": "w31161"
}

{
  "id": "w31161",
  "title": "Generative AI at Work",
  "version": "November 2023 working paper",
  "finding": "Following a staggered rollout of AI assistance, 5,179 customer-support agents resolved 14% more issues per hour on average when given access to the tool.",
  "url": "https://www.nber.org/papers/w31161"
}

search_series {
  "query": "nonfarm business output per hour quarterly index"
}

[
  {
    "id": "productivity",
    "title": "U.S. nonfarm business labor productivity"
  }
]

series_metadata {
  "series_id": "productivity"
}

{
  "id": "productivity",
  "title": "U.S. nonfarm business labor productivity",
  "measure": "Output per hour worked",
  "units": "Index, 2017 = 100",
  "frequency": "Quarterly",
  "seasonal_adjustment": "Seasonally adjusted",
  "vintage": "BLS third-quarter 2026 table",
  "url": "https://www.bls.gov/charts/productivity-and-costs/nonfarm-business-sector-indexes.htm"
}

series_observations {
  "series_id": "productivity",
  "start": "2022-10-01",
  "end": "2024-10-01"
}

{
  "observations": [
    {
      "date": "2022-10-01",
      "value": 110.219
    },
    {
      "date": "2024-10-01",
      "value": 116.978
    }
  ],
  "url": "https://www.bls.gov/charts/productivity-and-costs/nonfarm-business-sector-indexes.htm"
}

percent_change {
  "start_level": 110.219,
  "end_level": 116.978
}

{
  "percent_change": 6.132
}

The examples replay model decisions through an in-process MCP teaching server. The paper record is a short paraphrase. The two index levels come directly from the BLS third-quarter 2026 table. Live FRED observations can change with revisions. BLS source table

5 Check the claim against the source

Which version did the agent read? The NBER working paper reports a 14% average gain for a sample of 5,179 workers. The published article reports 15% for 5,172 workers. Both numbers have a source. A citation that identifies the version tells the reader which number belongs in the answer. Working paper · Published article

How much of the national change came from AI? The firm study follows a rollout. The national index compares output per hour on two dates. Attributing some of that change to AI requires a credible estimate of what national productivity would have been without adoption. Calling both tools gives the agent enough evidence to explain that limit.