Search across all documentation pages
The "best" agent framework on a blog post is irrelevant if your team cannot debug it under load. Skills, language fluency, and staffing depth are first-class selection criteria - equal to orchestration features.
Use this cheatsheet to match people to stacks without pretending everyone will retrain overnight.
| Team shape | Lean toward | Think twice about |
|---|---|---|
| Python backend / data platform | LangGraph, Pydantic AI, CrewAI, LlamaIndex, MS Agent Framework | TS-only stacks as the core runtime |
| TypeScript product eng | Vercel AI SDK + thin TS/Python worker | Python-only graphs with no TS owners |
| .NET / Microsoft enterprise | Microsoft Agent Framework | Niche Python-only tools without MS bridge |
| Tiny startup (2-5 eng) | Thin SDK or Pydantic AI; no-code for glue | Multi-framework estates |
| Platform team serving many squads | One shared host + adapters | Each squad invents a stack |
| Ops-heavy, few engineers | n8n/Zapier-class automation | Custom multi-agent graphs |
| Research-heavy ML group | LangGraph / LlamaIndex | No-code as core experiment harness |
| Compliance-heavy enterprise | Frameworks with audit story and strong platform eng | Demo-oriented crew toys without ops plan |
Rate each 0-2 (none / some / strong). Totals guide shortlist weight.
Reading the inventory (rule of thumb):
| Pattern | Interpretation |
|---|---|
| High Python + workflows | LangGraph / MS workflows comfortable |
| High Python + weak workflows | Start Pydantic AI or native SDK; add graphs later |
| High TS UI + medium backend | Vercel AI SDK front; backend host owned by backend cohort |
| Low LLM eval skills | Pick simpler topology; budget training before multi-agent |
| High ops, low eng headcount | No-code + few critical code paths |
| High systems, low LLM | Custom thin loop possible; still learn stop/eval discipline |
| Framework / approach | Skills that make it thrive | Skills gap failure mode |
|---|---|---|
| LangGraph | Graph thinking, Python, state/debug discipline | Over-complex graphs nobody can reason about |
| CrewAI | Product storytelling via roles, Python scripting | Crews that never converge; weak debugging culture |
| Microsoft Agent Framework | Enterprise integration, MS ecosystem fluency | Fighting cloud/identity assumptions |
| LlamaIndex | Retrieval, chunking, data quality habits | "Agent" wrapper hides bad retrieval metrics |
| Pydantic AI | Typed Python, schema design, service boundaries | Expecting full multi-agent platform out of the box |
| OpenAI Agents SDK (native) | Comfort with provider docs, fast iteration | Surprise lock-in; multi-cloud strategy absent |
| Vercel AI SDK | TS/React product craft, streaming UX | Secrets and long tools stuck in the edge layer |
| No-code builders | Process mapping, integration admin skills | Unowned sprawl; no tests for critical branches |
| Custom runtime | Strong platform eng + LLM ops ownership | Abandoned loop after author leaves |
| Headcount owning the agent | Practical default | Anti-default |
|---|---|---|
| 1 engineer | Thin library + ruthless scope | Multi-agent platform tour |
| 2-3 engineers | One backend host + clear UI split | Three frameworks "for flexibility" |
| One squad (5-8) | Shared patterns repo + one primary orchestrator | Each feature picks a new stack |
| Multi-squad platform | Internally supported runtime/adapters | Uncoordinated DIY everywhere |
| Enterprise CoE | Approved shortlist + ADR gates | Shadow IT no-code with production secrets |
| Move | Typical training load | Notes |
|---|---|---|
| Python API team → Pydantic AI | Low | Types already familiar |
| Python API team → LangGraph | Medium | Graph/state concepts |
| Python API team → CrewAI | Low-medium | Easy start; harder production debugging |
| TS app team → Vercel AI SDK | Low | Fits existing mental models |
| TS app team → Python LangGraph core | High | Split ownership or retrain |
| Ops team → n8n agents | Low | Governance still required |
| Any team → custom runtime | Medium-high | You must teach loop + ops standards |
| Any team → multi-agent first | High | Coordination skills rare |
If training load is high and deadline is near, change the problem shape (simpler architecture) before forcing a heroic framework adoption.
| Role | Why framework choice cares |
|---|---|
| Agent runtime owner | Without one, every framework rots |
| Tool/API owner | Side effects dominate failures |
| Eval owner | Prevents silent quality drift |
| Security reviewer | Tool allowlists and sandboxes |
| Product engineer (UI) | Streaming and approval UX |
| Platform/SRE | Budgets, rollbacks, capacity |
A framework with perfect features still fails if three of these roles are "nobody."
If you cannot check the first two boxes, simplify architecture before expanding framework scope.
| Anti-pattern | Why it fails | Better move |
|---|---|---|
| Picking based on Twitter hype | Skills mismatch | Spike with your team, not influencers |
| Forcing Python graphs on a TS org | Bus factor = 1 contractor | UI TS + backend hire/train plan |
| "We'll hire our way out" | Hiring markets lag roadmaps | Choose operable defaults now |
| No-code with eng-only ownership guilt | Neither side owns quality | Assign admins + review gates |
| Custom runtime as resume-driven design | Maintenance orphan risk | Require platform sponsor |
| One genius owns Crew + Graph + RAG | Single point of failure | Standardize and document |
For production ownership, yes when gaps are large. A slightly "worse" stack you can operate beats a perfect stack that only one person understands.
Yes if you fund training, pair programming, and a non-critical first pilot. Do not stretch on the revenue-critical path without a mentor.
Define contracts (schemas, events) and ownership boundaries. Common pattern: Vercel AI SDK UI, Python worker for tools and durable flow.
Often for demos and role metaphors. Production debugging may still demand senior help when crews loop or drift.
When your engineers already live in Microsoft identity, telemetry, and integration patterns, and enterprise connectors dominate the work.
No. You need software engineers comfortable with state machines, tools, and evals. ML specialists help with model selection and RAG quality, not only graph wiring.
Attach skills inventory scores to the framework ADR. If scores are low, the plan must include training time or a simpler architecture.
Budget a migration or formal knowledge transfer before production traffic. Contractor-favorite frameworks are a classic lock-in vector.
They are real operational skills. Treat them with the same access control, review, and testing expectations you apply to code when impact is high.
Small teams should default single-agent. Multi-agent multiplies coordination and ownership surfaces faster than headcount usually allows.
Mandate interfaces, observability fields, and an approved shortlist. A single framework can be the default without being a religious monopoly.
Often Pydantic AI or a native SDK with strict stop budgets - add LangGraph when workflow complexity appears.
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.
Reviewed by Chris St. John·Last updated Jul 16, 2026