Dev Tools

Spoiler: I deleted LangChain from a production project last month and the codebase got shorter.

The Setup

LangChain has 47 ways to do one thing and the docs keep deprecating themselves. Once you write a non-trivial agent, you're reading the source anyway. So skip the middle step — write the source yourself, in 200 lines, with types you understand.

{`pip install litellm pydantic
# That's it. That's the dependency list.`}

The Money Pattern

A typed client, a chain runner, and a tool dispatcher. Three primitives, zero magic. You can read the whole thing in a coffee break and your IDE actually knows what's going on.

{`from pydantic import BaseModel
from litellm import completion

class Step(BaseModel):
    role: str
    content: str

class Chain:
    def __init__(self, model: str, system: str):
        self.model = model
        self.history: list[Step] = [Step(role="system", content=system)]

    def run(self, user: str) -> str:
        self.history.append(Step(role="user", content=user))
        res = completion(
            model=self.model,
            messages=[s.model_dump() for s in self.history],
        )
        out = res.choices[0].message.content
        self.history.append(Step(role="assistant", content=out))
        return out

bot = Chain("anthropic/claude-sonnet-4-5", "You triage hail claims.")
print(bot.run("Claim 4821: roof, two skylights cracked."))
print(bot.run("What did I just tell you?"))`}

The Catch

You lose the integrations. No prebuilt 47-class retriever, no off-the-shelf SQL agent, no vector store adapter zoo. For 80% of real products, those abstractions were costing you more than they saved. For the other 20%, fine, use the framework — just know which one you are.

The Verdict

Frameworks are great until they aren't, and LangChain hit "aren't" two years ago. Write the 200 lines. Own the abstraction. Sleep better.