Framework Comparison Matrix: LangGraph, CrewAI, Microsoft Agent Framework & More
Use this page when you need a ranked, side-by-side reference - not a vendor brochure. Scores are engineering fit heuristics for common production agent work in 2025-2026, not absolute quality rankings.
Version pins move; verify APIs and maturity claims at build time.
How to Use This List
- Start with Quick pick by job, then confirm in the master matrix.
- Prefer the simplest stack that meets orchestration, language, and ops needs.
- Pair with What Actually Differs Between Agent Frameworks for axis definitions.
- Re-score after a 1-2 day spike on your real tools, not a hello-world chat.
Quick Pick by Job
| If your primary job is… | First shortlist | Usually skip first |
|---|---|---|
| Explicit branching, resume, HITL graphs | LangGraph; Microsoft Agent Framework workflows | Pure no-code; crew-only demos |
| Multi-role research / content crews fast | CrewAI | Heavy enterprise graph setup |
| Enterprise connectors + governance gravity | Microsoft Agent Framework | Greenfield TS-only startups with no MS stack |
| RAG / document agents first | LlamaIndex (± separate orchestrator) | Crew frameworks without a retrieval plan |
| Typed Python tools + structured outputs | Pydantic AI | Visual builders for core logic |
| Single-provider agent primitives quickly | OpenAI Agents SDK (or peer native kits) | Multi-cloud portability-first designs |
| Streaming web UX + tool UI in TypeScript | Vercel AI SDK | Using it alone as durable workflow engine |
| Ops automation owned by non-dev teams | n8n / Zapier-class no-code | Complex regulated multi-agent systems |
| Maximum control, minimal dependencies | Custom runtime | Teams that cannot staff loop + ops ownership |
Master Comparison Matrix
Ratings: H = strong default fit, M = workable with effort, L = weak default fit. "Maturity" means production-operability signal for agent control planes, not model quality.
| Framework / approach | Orchestration style | Primary languages | Multi-agent bias | State / resume | Observability path | Ecosystem gravity | Maturity signal | Honest downside |
|---|---|---|---|---|---|---|---|---|
| LangGraph | Graph / state machine | Python, JS | Bring-your-own topology via nodes | H checkpoints & shared state | LangSmith + self-host patterns | LangChain ecosystem | H (1.0+ line) | Concept weight; easy to over-graph |
| CrewAI | Roles, tasks, crews | Python | Sequential / hierarchical crews | M task outputs & memory plugins | Platform/logging varies by setup | Fast multi-agent DX | H for crew-style apps (1.14+) | Convergence debugging; less rigid edges |
| Microsoft Agent Framework | Agents + graph workflows | Python, .NET (verify) | Conversational multi-agent + workflows | H workflow-oriented | Azure-friendly ops story | Semantic Kernel / enterprise | H (1.0) | Stack gravity; learning Autogen/SK history |
| LlamaIndex | Query engines + agent wrappers | Python, TS | Router / tool agents around data | M (session + index state) | Callbacks / integrations | RAG ecosystem depth | H for retrieval apps | Orchestration depth secondary to data layer |
| Pydantic AI | Schema-first agent | Python | Mostly single-agent composition | M (app-owned persistence) | Logfire / OTel-style options | Pydantic/type-safety culture | H for typed backends | Not a full enterprise workflow suite |
| OpenAI Agents SDK (native class) | Agent + tools + handoffs | Python (+ others per SDK) | Handoffs & guardrails | M (patterns evolving) | Provider tracing tools | OpenAI platform | M-H (fast-moving) | Provider coupling; multi-cloud cost |
| Vercel AI SDK | App SDK + agent loop helpers | TypeScript | App-level patterns | L-M (often externalize) | App logs + platform | Next.js / Vercel deploy | H for AI product UX (v6) | Not a substitute for durable backend graphs |
| No-code (n8n, Zapier AI, …) | Visual workflows | N/A (UI) | Linear / light branching | M node data | Platform run history | SaaS integrations | M for ops automation | Testing, complex logic, review friction |
| Custom runtime | Whatever you build | Your stack | Whatever you design | H if you invest | DIY / OTel | None (full ownership) | Varies by team | You own every missing feature forever |
Ranked Defaults by Constraint
1. Need durable multi-step control with audits
- LangGraph - explicit state, edges, checkpoints.
- Microsoft Agent Framework - workflows + enterprise controls.
- Custom runtime - if frameworks fight compliance constraints.
- CrewAI / thin native SDKs - only if you add substantial host discipline.
2. Need fastest multi-role prototype in Python
- CrewAI - roles/tasks map to stakeholder language.
- LangGraph - if you already think in graphs.
- OpenAI Agents SDK - handoffs without crew metaphors.
- No-code - only for shallow role cosplay over SaaS steps.
3. Need RAG quality more than choreography
- LlamaIndex - indexes, routers, evals around retrieval.
- Pydantic AI / custom + your vector layer - when retrieval is thin.
- LangGraph wrapping retrieval nodes - when the workflow is the product.
- CrewAI - viable, but do not confuse role prompts with retrieval engineering.
4. Need typed backend correctness
- Pydantic AI - models, tools, validated outputs.
- Custom FastAPI/Express loop - full control.
- LangGraph with typed state - more moving parts.
- No-code - weak for strict schema contracts as core logic.
5. Need polished streaming UI in TypeScript
- Vercel AI SDK on the client/edge surface.
- Backend Pydantic AI / LangGraph / custom for durable work.
- Native provider web SDKs - if single-provider and simple.
- Python-only stacks - add a TS bridge rather than forcing Python into the browser.
6. Need non-engineer-owned automation
- n8n / Zapier-class tools.
- Thin scripts behind a form UI.
- Full agent frameworks - usually overkill and under-owned.
- Custom multi-agent graphs - last resort for ops teams.
Language and Hiring Fit (compressed)
| Team reality | Lean toward |
|---|---|
| Python data/ML platform team | LangGraph, CrewAI, LlamaIndex, Pydantic AI, MS Agent Framework |
| TypeScript product team | Vercel AI SDK + thin backend; LlamaIndex.TS when RAG-heavy |
| .NET / Microsoft enterprise | Microsoft Agent Framework first pass |
| Mixed, multi-cloud | Portable tools + OpenRouter-style routing; avoid single-provider agent core |
| Tiny team, many SaaS apps | No-code first, code the exceptions |
See Matching Team Skill Set to Framework Choice for a fuller checklist.
Feature Myths vs Reality
| Claim you will hear | More accurate reading |
|---|---|
| "Supports multi-agent" | Check topology: crew vs handoff vs graph join |
| "Has memory" | Run transcript ≠ long-term memory store design |
| "Production ready" | Ask about timeouts, auth, evals, and on-call docs |
| "Model agnostic" | Often true for chat completions; weaker for provider-only agent APIs |
| "Enterprise grade" | Translate to: SSO, RBAC, audit logs, VPC, support SLAs |
| "No lock-in" | Measure: tool wrappers, state format, prompt DSL, hosting |
Anti-Patterns When Using This Matrix
| Anti-pattern | Why it fails | Better move |
|---|---|---|
| Picking the highest "maturity" cell always | Ignores task shape and team skills | Weight columns for your constraints |
| Ranking from Twitter demos | Demo != resume, HITL, or cost control | Spike with your tools and stop rules |
| One framework for UI + batch + ops | Wrong abstraction per layer | Compose specialized layers |
| Ignoring no-code out of pride | Burns eng time on glue | Use visual builders for glue workflows |
| Ignoring custom out of fear | Frameworks can be the wrong dependency | Prototype a 100-line loop for simple agents |
Default Decision Order
- Is this even an agent? Fixed scripts and single tool calls win when enough.
- What orchestration shape is mandatory? Graph, crew, thin loop, or visual workflow.
- What language ships the product? Do not adopt a stack your team cannot debug.
- What must be portable? Tools, schemas, prompts, model routing.
- What ops hooks are non-negotiable? Tracing, auth, budgets, eval gates.
- Spike two finalists for 1-2 days. Keep scorecards; write an ADR.
FAQs
Is LangGraph always the "most production" choice?
No. It is a strong default for explicit control and resume, but a typed single-agent service or no-code workflow may be more production-appropriate for simpler jobs.
Why is CrewAI not ranked first for production graphs?
CrewAI optimizes role/task multi-agent authoring. Production graph problems (strict edges, complex joins, checkpoint semantics) often fit LangGraph or Microsoft workflows better.
Can I use LlamaIndex and LangGraph together?
Yes. Common pattern: LlamaIndex for retrieval quality, LangGraph for multi-step control. Keep interfaces clean so neither library owns everything.
Where does Pydantic AI beat heavier frameworks?
Backend services that need validated structured outputs, typed tools, and clear Python ergonomics without multi-agent ceremony.
Are native SDKs "cheating" on portability?
They are a conscious bet. They are honest when your provider strategy is stable; expensive when you must multi-home models and agent APIs later.
Is Vercel AI SDK comparable to LangGraph?
Only partially. AI SDK excels at TS application UX and streaming. LangGraph excels at durable orchestration. Many products use both at different layers.
When does Microsoft Agent Framework outrank LangGraph?
When enterprise connectors, governance expectations, and Microsoft platform alignment reduce total cost of ownership more than graph-ecosystem familiarity does.
Should startups default to no-code agents?
For internal automation, often yes. For customer-facing core product agents, code with tests and reviews usually wins sooner than people expect.
How do I score "maturity" in my own matrix?
Use: release stability, debug tooling, clear stop/timeout controls, auth patterns, community answers to production incidents, and your team's prior art.
What two finalists should most Python teams spike?
Often one graph-oriented (LangGraph or MS workflows) and one thin/typed (Pydantic AI or native SDK) - unless the problem is clearly crew-shaped or RAG-shaped.
How often should we revisit the matrix decision?
At major roadmap shifts: multi-agent needs, new language ownership, compliance gates, or repeated production incidents tied to control-flow limits.
Do OpenRouter or model gateways change the matrix?
They mainly affect model portability and cost routing, not orchestration style. Still valuable: keep model access behind an interface so framework choice stays separable.
Related
- What Actually Differs Between Agent Frameworks - axes behind the table
- Choosing a Framework Basics - same-task dual spike
- Matching Team Skill Set to Framework Choice - skills and language fit
- Framework-Free vs Framework-Based: When to Roll Your Own - custom column explained
- Choosing a Framework Best Practices - evaluation discipline
- When Visual Workflow Tools Beat Writing Custom Agent Code - no-code path
Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana
@solana/web3.js+@solana/spl-token.