Back office ops · Production

Madrigal Pharmaceuticals builds a pharmaceutical-grade multi-agent AI research platform with LangChain and LangSmith

The problem

Madrigal's enterprise data was fragmented across structured systems, unstructured documents, external sources, and real-time APIs with no unified way to search, analyze, and synthesize information at scale; each source behaved differently with incompatible formats, access patterns, and expectations.

Workflow diagram · grounded in source

Employee submits query

trigger

“Madrigal employees to be able to easily search, analyze, and synthesize relevant data spread across our enterprise within appropriate access controls”

Orchestrator routes to agents

routing

“The orchestrator is where intent meets execution. It receives a task and decides what needs to happen next: which capabilities are required, which agents should run, what should happen in parallel and when to bring everything back together.”

Parallel agents query data

ai_action

“Multiple agents run in parallel, each focusing on a different slice of the problem. By the time the system comes back together, it combines fully formed pieces of work. The orchestration agent can divide a research question into three pa…”

Results stored in shared workspace

integration

“a place where every agent can write what it finds and read what others have already done. This becomes the system's memory. Every result is stored. Every source is tracked. Every intermediate step is available for reuse.”

Cited answers delivered

output

“ensuring every response was clearly cited”

Production failures improve evals

feedback_loop

“production failures feed back into our LangSmith datasets automatically. Every meaningful error becomes a new test case. The eval suite grows from real failures, not synthetic scenarios.”

Reported outcome

Madrigal's multi-agent platform enables new use cases to go from weeks of development to hours, with prototype-to-enterprise deployment taking weeks rather than months, and the system scales to new domains without rewriting orchestration logic or adding new infrastructure.

Reported metrics

New use case development timeweeks is reduced to hours

Prototype to enterprise deployment timeweeks, not the months we'd budgeted

Enterprise system build and launch timeweeks, not months

Reported stack

LangChainLangGraphLangSmithDeepAgentsGitHub

Source

https://www.langchain.com/blog/customers-madrigal

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Madrigal's multi-agent platform enables new use cases to go from weeks of development to hours, with prototype-to-enterprise deployment taking weeks rather than months, and the system scales to new domains without rew…

What tools did this team use?

LangChain, LangGraph, LangSmith, DeepAgents, GitHub.

What results were reported?

New use case development time: weeks is reduced to hours; Prototype to enterprise deployment time: weeks, not the months we'd budgeted; Enterprise system build and launch time: weeks, not months (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Employee submits query → Orchestrator routes to agents → Parallel agents query data → Results stored in shared workspace → Cited answers delivered → Production failures improve evals.