What is this for? To compare popular frameworks for building agent and RAG apps—and explain when LangGraph fits vs plain LangChain.
Why does it exist? Frameworks speed up prototyping, but they do not remove the need for evals, guardrails, or clear ownership of side effects. Pick the tool that matches your workflow shape.
Four names show up constantly:
| Framework | Plain-English strength | Typical use |
|---|---|---|
| LangChain | Chains, tools, prompt pipelines | General LLM apps |
| LangGraph | Cyclic, stateful graph control flow | Reasoning loops, multi-agent graphs |
| LlamaIndex | Data connectors + indexing | Knowledge / RAG assistants |
| AutoGen | Agent conversation patterns | Multi-agent message passing |
| CrewAI | Role-based orchestration | Task teams with shared goals |
Think "batteries," not "brain." Your product still needs schemas, threat models, and CI gates.
| LangChain | LangGraph | |
|---|---|---|
| Plain-English idea | Component framework | Graph that can loop and remember |
| Flow shape | Mostly linear chain | Cyclic graph with state |
| Good for | Sequential tasks | Reasoning loops, branches, HITL pauses |
| Analogy | Straight line | Flowchart |
Broad ecosystem: prompt templates, retrievers, tool wrappers, runnables, many vendor integrations. Good when you want one toolkit from prototype to agents.
Risk: deep abstraction stacks that are hard to debug—keep traces on and avoid nesting magic.
Graph-based framework for looping, stateful agent workflows:
compile() = turns the graph definition into a runnable, checkpointable application.Example shape:
graph = StateGraph(AgentState)
graph.add_node('triage_agent', triage_node)
graph.add_node('infra_agent', infra_node)
graph.add_conditional_edges('triage_agent', router_function)
graph.add_edge('infra_agent', 'triage_agent')
app = graph.compile()Use LangGraph when the workflow needs loops, branches, human approval checkpoints, or recursion limits.
Centered on connecting data → indexes → query engines. Strong default for RAG-heavy assistants (files, Notion, databases). You still own chunking strategy, evaluation, and citation policy.
Patterns for multiple agents exchanging messages (assistant/user proxy, group chat). Useful when the problem naturally looks like a conversation among roles. Watch cost: chatty agents amplify tokens; impose turn limits and termination conditions.
Role/goal/backstory style crews that divide labor on a task. Friendly mental model for "researcher + writer + reviewer." Same multi-agent cautions: typed artifacts beat free-form chatter; measure whether the crew beats a single agent.
| If your problem looks like... | Lean toward |
|---|---|
| RAG-heavy knowledge app | LlamaIndex or LangChain retrievers |
| Tool-using single agent | LangChain or a thin custom loop |
| Cyclic loops, HITL, multi-agent graphs | LangGraph |
| Explicit multi-agent chat | AutoGen or CrewAI (on your eval suite) |
Team constraints—observability, deploy story, license, who already knows the stack—often dominate micro-benchmarks.
Regardless of framework, own these in your code:
If a framework makes those hard, wrap it thinly or skip it.
A framework-agnostic skeleton—swap the complete and retrieve adapters for LangChain/LlamaIndex later.
from dataclasses import dataclass
@dataclass
class FrameworkAdapters:
retrieve: callable # query -> list[str]
complete: callable # prompt -> str
def answer_with_rag(q: str, fx: FrameworkAdapters) -> str:
chunks = fx.retrieve(q)[:4]
context = "\n---\n".join(chunks)
prompt = (
"Answer using only CONTEXT. If missing, say you do not know.\n"
f"CONTEXT:\n{context}\n\nQUESTION: {q}"
)
return fx.complete(prompt)
fx = FrameworkAdapters(
retrieve=lambda q: [f"Policy: refunds within 30 days for '{q}'"],
complete=lambda p: "Refunds are available within 30 days."
if "30 days" in p else "I do not know.",
)
print(answer_with_rag("refund window?", fx))Learn the framework's tracing first; demos without traces become production mysteries.
Spike two candidates for 48 hours on the same golden set and the same tools. Score quality, p95 latency, cost, and "time for a new engineer to find the prompt."
Pin versions, snapshot prompts used in CI, and treat framework upgrades like model upgrades: run the full suite before merging.
For a single tool-using agent, a 100-line host loop plus your vendor's tool-calling API is often enough. Reach for LangChain/LlamaIndex when connectors and retrievers would otherwise dominate your calendar. Reach for LangGraph when you need loops, branches, or HITL. Reach for AutoGen/CrewAI when you have measured a multi-agent win.
Keep golden cases and graders outside the framework's preferred storage format when you can—plain JSON/YAML in git. Frameworks should be replaceable adapters under a stable evaluation harness.
Use LangChain, LlamaIndex, AutoGen, or CrewAI as accelerators—but use LangGraph when you need cyclic stateful control; keep validation, evals, and safety in your own thin layer.