I tried CrewAI and it feels too fragile for production

That’s a fair take. CrewAI can be great for prototypes and demos, but in production the weak spots usually show up fast: orchestration brittleness, state handling, retries, observability, and hard-to-debug agent handoffs.

If you want something more production-friendly, it helps to separate the problem into layers:

What usually breaks in “agentic” systems

• Unclear control flow: agents deciding too much at runtime
• Hidden state: prompts and outputs becoming the de facto database
• Poor retries/timeouts: one bad model call cascading into failure
• No tracing/evals: hard to know why something failed
• Tool coupling: brittle dependencies on prompt format or tool output shape

More robust alternatives

Depending on what you need:

• LangGraph
  Better for explicit, stateful workflows with graphs, retries, and checkpointing. Often a stronger choice if you want agent-like behavior but more control.

• Temporal / Durable workflows
  If reliability is the priority, model calls can be one step in a workflow engine. Very production-oriented for retries, timers, resumability, and auditability.

• Autogen / custom orchestrator
  Can work, but I’d usually recommend a thin custom framework only if your use case is simple and you can keep control flow explicit.

• Non-agentic pipeline + tool use
  Often the most stable approach: deterministic workflow first, LLM only where needed.

A production pattern that tends to work

• Keep orchestration deterministic
• Use the LLM for classification, extraction, synthesis
• Put tools behind strict schemas
• Add timeouts, retries, idempotency
• Log every step with structured traces
• Build evals for the failure modes you care about

Rule of thumb

If the system must be:

• auditable
• reliable
• recoverable
• repeatable

then agent frameworks should usually be a thin layer, not the core control plane.

If you want, I can suggest:

1. a CrewAI replacement stack for production, or
2. a reference architecture for building reliable LLM workflows.