llm-frameworks

name

llm-frameworks

description

Use this skill when choosing between LangChain, LangGraph, CrewAI, Semantic Kernel, or custom implementation, when someone asks 'which framework should we use?', 'should we use LangGraph or LangChain?', 'is CrewAI a good fit?', 'should we build custom?', 'we need a framework for our agents', 'comparing LLM frameworks', or when evaluating framework trade-offs for agent orchestration. Also trigger for topics: LCEL, StateGraph, Crew, Semantic Kernel skills, framework migration.

LLM & Agent Framework Selection

Framework Comparison

Aspect	LangChain	LangGraph	CrewAI	Semantic Kernel	Custom
Loops / cycles	Manual	Native	Manual	Manual	Full control
Complex branching	Limited	Excellent	Limited	Limited	Full control
Multi-agent	Basic	Complete	Main focus	Basic	Full control
TypeScript support	Yes	Yes	No	Yes	Yes
Ready integrations	100+	Inherits LC	Limited	Microsoft stack	None
Human-in-the-loop	Manual	Native (interrupt)	Manual	Manual	Full control
Workflow visualization	No	Mermaid	No	No	Custom
Performance overhead	Medium	Medium	Low-medium	Medium	None

Decision Tree

Workflow has loops, cycles, or complex conditional branching?
├── YES → LangGraph
└── NO
    ├── Multiple agents with coordination?
    │   ├── TypeScript required → LangGraph
    │   └── Python acceptable
    │       ├── Speed over control → CrewAI
    │       └── Control over speed → LangGraph
    ├── Linear RAG pipeline, no agent coordination → LangChain
    ├── Microsoft / .NET / Azure stack → Semantic Kernel
    └── Compliance forbids external deps / performance critical / <50% fit → Custom

When Each Framework Wins

Use LangChain when:

• Pipeline is linear or only slightly conditional
• 100+ integrations are valuable (databases, vector stores, LLMs)
• Rapid prototyping is the priority
• Team already knows LangChain

Use LangGraph when:

• Workflow has loops (draft → critique → revise → repeat)
• Human review and approval at specific steps is required
• Need full visualization of the state machine
• Shared state between agents across multiple steps

Use CrewAI when:

• The "team with roles" mental model fits the domain naturally
• Python-only is acceptable
• Speed of development outweighs fine-grained control
• Role-based agent delegation is the primary pattern

Use Custom when:

• Compliance prohibits third-party dependencies
• Performance is critical and framework overhead is measurable
• Framework covers < 50% of actual use cases
• Full auditability of every operation is required

Trade-off Summary

• LangGraph vs LangChain: LangGraph adds native cycle support and state management at the cost of slightly more setup. For any workflow with loops, LangGraph is strictly better.
• LangGraph vs CrewAI: LangGraph gives more control; CrewAI abstracts coordination. CrewAI is faster to prototype if the role-based model fits; harder to customize when it doesn't.
• Framework vs Custom: frameworks reduce initial development time by ~60%; custom implementations reduce long-term maintenance burden and eliminate version dependency issues. The break-even is typically at 12–18 months.

Migration Considerations

Migrating between frameworks is expensive. The main costs:

• Rewriting state management (LangChain chains → LangGraph StateGraph)
• Re-implementing tool integrations
• Re-testing all workflows

Establish the workflow complexity early. If loops are anticipated within 6 months, start with LangGraph even for simpler initial cases.

Perguntas diagnósticas

1. Does the workflow have cycles (output feeds back as input to an earlier step)?
2. Is human approval required at any point in the workflow?
3. Is TypeScript a hard requirement?
4. Does the team have existing familiarity with any of these frameworks?
5. Are there compliance restrictions on third-party dependencies?
6. What percentage of the actual use cases does each candidate framework cover?