Most agent applications return a completed answer: the request runs to completion, then the full response arrives at once. The alternative interleaves retrieval and generation, streaming tokens to the client while the agent is still searching and reading. In late 2025 I ran a contained experiment to understand the streaming mode end to end, and Lang Stream is the result: a full-stack agent application with a LangGraph agent, a FastAPI backend, and a Next.js frontend, containerized to start with one command. The first commit is from late October 2025 and the last from early November 2025. It was a study, not a product, and this writeup treats it as one.
Component choices
LangGraph was the reason the project existed. Chain-style agent frameworks make the default path simple and everything else opaque. LangGraph instead defines the agent as an explicit graph of states and transitions, so the control flow is inspectable: where a search result enters the flow and where generation resumes are visible in the graph definition rather than implicit in framework behavior. For an agent that interleaves searching, reading, and writing, that explicitness is the property that matters.
FastAPI was chosen for one dominant property: streaming is native to
it. An async Python framework provides token-level streaming over HTTP
directly, and it runs in the same language as the agent, so no
serialization boundary sits between the graph and the endpoint. The
resulting API surface is a single POST to /api/v1/agent with a query,
and the response streams back as the agent works.
The rest of the stack is standard by intent. PostgreSQL 16 for
persistence. Redis 7 as both cache and message broker. Next.js 15 with
shadcn/ui on the frontend, chosen to avoid hand-building accessible
components inside a two-week window. Every service runs in Docker on a
shared network with named volumes for persistent data, and
docker-compose up --build starts all four containers with hot
reloading on both the Python and TypeScript sides. Getting that
development loop right took real effort and paid for itself: the
environment rebuilds around code changes while the experiment is
running, which is the property a study project needs most.
Scope, stated precisely
Lang Stream is a scaffold. The README is setup instructions almost top to bottom: environment variables, ports, how to run the tests, how to add a shadcn component. That reflects the project’s actual job, which was to hold the architecture fixed so the agent graph and the streaming path could be studied in isolation.
The limits are explicit. There is one real endpoint. The frontend is a chat surface, not an application. The test suite is pytest, concentrated on the auth module, with coverage consistent with a two-week project. The README does not present the system as deployable, and it is not.
What the experiment established
The results split into two categories. The first is mechanics: how a LangGraph agent hands tokens to an async endpoint, where Redis belongs when responses are produced and consumed at different rates, and how to keep a mixed Python and TypeScript stack workable inside Docker.
The second is architectural, and more useful: streaming is a user experience decision with consequences at every layer, and it cannot be added at the end. The graph, the HTTP framework, and the frontend rendering all have to pass partial results through, and choosing a component that blocks (a synchronous framework, a request-response-only client) forecloses the design before any streaming code is written.
Archived, deliberately
The repo’s last commit is from November 9, 2025, and the repository is archived rather than left nominally active on my profile. The experiment completed its purpose in two weeks. The LLM and streaming plumbing in Vidura was built afterward on top of what this project established, at far lower cost than learning the same lessons in a production system.
The code remains public under MIT as a working reference: LangGraph,
FastAPI, Next.js, PostgreSQL, and Redis wired into one deployable unit
with the streaming path intact, reproducible with a single
docker-compose up.