Cellaflow
https://www.cellaflow.com/ Cellaflow — Safe retries for AI agent tool calls. The question I'm interested in is: how do you make sure recovery of an AI agent doesn't repeat the side effect? Recovering an agent's state after a crash is the easy half, and most frameworks already have a checkpointer for it. The hard half is recovering after the agent has already done something irreversible. Sometimes the downstream API has idempotency built in. But often you don't control the API, the operation isn't naturally idempotent, or you need to coordinate several side effects. Sending an email, creating a ticket, triggering a webhook, provisioning something, or mutating another system can all leave you with the same problem. Cellaflow focuses on the side effect, rather than just the agent's state: which agent is allowed to perform an operation, what's already been done, and how to safely recover when the agent crashes or disappears mid-execution. It is a runtime that sits underneath your existing agent framework, so you can use it with LangGraph, CrewAI, AutoGen, or your own agent code without migrating the agent itself. Still early. I'd like to talk to anyone running agents in production who's hit duplicate tool calls, duplicate side effects, or a workflow that quietly never finished.