I worked with CrewAI for years. Here's why I now build on Google and Claude.
CrewAI, LangGraph and AutoGen taught me a lot — especially about prototyping multi-agent systems fast. I've since deliberately narrowed to two platforms I can run deeply in production: Google and Claude. This page is the honest story of that choice.
Matteo Gazzurelli · Fractional CTO & Agentic AI Consultant · Brescia, Italy
Google Partner (through MaGa Srl) — Google Cloud Select Co-sell Partner and Google Workspace Select Co-sell & Services Partner. Verify on the Google Cloud Partner Directory.
We'll figure out together whether CrewAI, Google, Claude — or none of them — fits your case. No sales pitch.
What is CrewAI — and why I now build on Google and Claude
CrewAI is an open-source Python framework for orchestrating role-based multi-agent systems, well-suited to prototyping agent ideas quickly. It's a capable tool — but for systems that have to survive production I've deliberately narrowed to two platforms I can run deeply: Google (Gemini Enterprise Agent Platform + ADK) and Claude (Agent SDK, Code), where identity, governance and observability are first-class.
That's a focus decision — depth over breadth — not a verdict that CrewAI is worse. If your team is already on CrewAI, I'm happy to help you get more out of it, or to plan a migration only if and when something needs to go to production.
What CrewAI taught me
For a couple of years I worked hands-on with CrewAI, alongside LangGraph and AutoGen. CrewAI's role-based orchestration model made it genuinely pleasant to express multi-agent collaboration — and, in my experience, it's often one of the fastest ways to get a sketch of an agentic system in front of a stakeholder. I don't regret a minute of that work.
Great for prototyping
Role-based crews, sequential and hierarchical flows — CrewAI let me explore ideas quickly and cheaply, which is exactly what early-stage work needs.
A real model of agents
Working across CrewAI, LangGraph and AutoGen taught me how these systems actually behave — where orchestration helps and where it quietly adds fragility.
Then production happened
Once systems had to survive contact with production — identity, governance, observability, long-term support — I found I wanted fewer moving parts and deeper platform support. That's what pushed the narrowing.
Where I focus now
Two platforms I can run deeply in production — Google and Claude — with framework prototyping (CrewAI included) still very much part of the early work.
Discovery & Strategy
- ✓Mapping the workflow and where agents genuinely help
- ✓An honest read on whether AI fits — sometimes it doesn't
- ✓Platform fit: Google, Claude, or a quick CrewAI prototype first
- ✓Where this typically lands as a 2–4 week engagement
Building on Google
- ✓Gemini Enterprise Agent Platform (formerly Vertex AI)
- ✓Agent Development Kit (ADK) for code-first agents
- ✓Identity, governance and the GCP ecosystem
- ✓A strong default when enterprise governance matters
Building on Claude
- ✓Claude Agent SDK for production deployment
- ✓Claude Code for the development workflow
- ✓Long context and strong instruction-following
- ✓A strong choice when the model itself is the differentiator
Prototyping & framework advice
- ✓Quick CrewAI / LangGraph prototypes to test an idea
- ✓Honest tradeoff guidance across frameworks
- ✓A migration path when a prototype needs to go to production
- ✓Team enablement and knowledge transfer
How this tends to play out
A few illustrative shapes of agentic work — a rough illustration, not a promise. The pattern I keep seeing: prototype fast with a framework, then run it deeply on one platform.
Customer support triage
The situation
A support inbox where most tickets follow a handful of patterns.
Prototype (CrewAI)
A CrewAI crew — triage → research → draft — is a quick way to test the idea.
Production (Google / Claude)
In production I'd typically run this on Google or Claude for identity, guardrails and observability.
Content operations
The situation
Recurring content across channels that needs consistency and a human check.
Prototype (CrewAI)
A research + writing + review flow prototypes nicely in a multi-agent framework.
Production (Google / Claude)
For ongoing use I lean on Claude's long context, or Google when it sits inside GCP.
Lead research & enrichment
The situation
Manual prospect research that's slow and inconsistent across the team.
Prototype (CrewAI)
Enrichment and qualification map cleanly onto role-based crews for a first pass.
Production (Google / Claude)
In production I prioritise data governance and auditability — where the two platforms help.
The stack I run deeply
Depth over breadth. Knowing two platforms well — their real tradeoffs, where they crack and where they shine — tends to be worth more in production than knowing many frameworks shallowly.
Google — Gemini Enterprise Agent Platform
- •Agent Runtime & Memory Bank: the platform formerly known as Vertex AI, evolved for agents
- •Agent Development Kit (ADK): code-first agent building
- •Identity & governance: Agent Identity, Agent Gateway, GCP ecosystem
- •Interop: MCP and A2A for tools and agent-to-agent communication
Claude — Agent SDK and Code
- •Agent SDK: Anthropic's platform for production agents
- •Claude Code: the development workflow itself
- •Long context and strong instruction-following
- •A strong choice when the model is the differentiator, or outside GCP
Where CrewAI fits today
CrewAI (with LangGraph and AutoGen) is part of my background, and still useful — mostly for prototyping. When I prototype, I tend to value:
How I tend to engage
Three honest models. Concrete pricing comes after Discovery, not before — and the timeframes below are typical ranges, not promises.
Discovery & Strategy
- •Map the workflow and where agents genuinely help
- •An honest read on whether AI fits at all
- •Platform fit: Google, Claude, or a CrewAI prototype first
- •A concrete plan — and the scope to price the next phase
Build & Deploy
- •Build on Google or Claude for production
- •Tool and system integrations
- •Testing, guardrails, observability
- •Deployment, handover and documentation
Ongoing Advisory
- •Fractional CTO support as the system evolves
- •Iteration as usage and requirements change
- •Team enablement and knowledge transfer
- •A steady hand rather than a one-off drop
Three honest engagement models
Concrete pricing comes after a Discovery engagement, not before. The timeframes below are typical ranges, not fixed products.
Discovery & Strategy
Typically 2–4 weeks
- ✓Workflow mapping and opportunity analysis
- ✓An honest read on whether AI fits
- ✓Platform fit: Google, Claude, or CrewAI prototype
- ✓A plan concrete enough to price the next phase
Build & Deploy
Typically 8–16 weeks
- ✓Production build on Google or Claude
- ✓Tool and system integrations
- ✓Testing, guardrails and observability
- ✓Deployment, handover and documentation
Ongoing Advisory
Typically 3–12 months
- ✓Fractional CTO support over time
- ✓Iteration as requirements evolve
- ✓Team enablement and knowledge transfer
- ✓A steady hand, not a one-off drop
Every engagement starts with a conversation to check whether AI — and which platform — actually fits your case.
Frequently Asked Questions
I spent a couple of years working hands-on with CrewAI, alongside LangGraph and AutoGen. It's a capable framework, and I still reach for it when prototyping multi-agent ideas quickly. For production, though, I've deliberately narrowed to two platforms I can run deeply: Google (Gemini Enterprise Agent Platform + ADK) and Claude (Agent SDK, Code). So CrewAI is part of my background, not the stack I lead with today.
Let's figure out the right stack
If you're weighing CrewAIagainst Google or Claude — or you're not sure agents are the answer at all — let's talk it through. I'll give you an honest read, including when the answer is “not yet”.
Quick Booking
Book directly on calendar🌍 Location
Brescia, Italy — serving clients globally
Let's talk about your stack
Pick a time and we'll figure out, honestly, whether CrewAI, Google, Claude — or none of them — fits your case.
Last updated: May 2026