Why a Checklist?
Multi-agent systems (MAS) are more complex than traditional software. When looking for a service provider, asking about "AI experience" isn't enough. You need to ensure the provider can solve architecture problems before they arise.
The 5 Critical Questions for Your First Meeting
1. Architecture: Do you use frameworks like LangGraph or AutoGen for orchestration?
Why it matters: Without an orchestration framework, your system will quickly become unmanageable. LangGraph and AutoGen are market leaders for multi-agent systems.
Red flag: "We write it ourselves" or "We just use OpenAI API directly"
Good answer: "We use LangGraph for deterministic workflows and AutoGen for collaborative agent teams."
2. Error Culture: How do you prevent infinite loops when an agent gets stuck?
Why it matters: A common problem in early agent projects is "infinite loops." An agent tries to solve a task, fails, and keeps trying—until the budget runs out.
Red flag: "That doesn't happen with us" or no concrete answer
Good answer: "We implement timeouts, retry limits, and supervisor agents that monitor the process."
3. Truth: Do you use "reviewer agents" to check for hallucinations?
Why it matters: The biggest fear about AI agents? That they invent facts. Professional providers implement reviewer agents—a second agent checks the first agent's output before it goes out.
Red flag: "LLMs don't hallucinate anymore" or "We use GPT-4, it's accurate enough"
Good answer: "We use self-correction patterns, grounding with RAG, and structured outputs."
4. Memory: How is the state of the conversation stored between agents?
Why it matters: Multi-agent systems must "remember" what has already happened. Without state management, agents lose context.
Red flag: "We store it in the session" or no clear answer
Good answer: "We use Redis for short-term state and PostgreSQL for long-term conversation history."
5. Security: How do you prevent agents from accessing database fields they shouldn't see?
Why it matters: When an AI agent gets access to your knowledge database (RAG), there's a risk: Should the agent tell an intern the CEO's salary?
Red flag: "We filter that in the prompt" or "That's not a problem"
Good answer: "We implement Access Control Lists (ACL) directly in the vector database and use RBAC (Role-Based Access Control)."
Bonus Question: Do you have references for production systems?
Ask for concrete examples of systems running in production—not just prototypes or demos. A good provider can name at least 2-3 references.
Conclusion: Save Expensive Lessons
Only providers who can give technically sound answers here can build a system that goes beyond a prototype. We've vetted providers in DACH and Europe who meet exactly these standards.