Anyone else hitting a wall with the "Day 2" side of shipping AI agents?

For the last six months, our team was completely focused on agent logic, testing prompts, trying out different multi-agent frameworks and building nice flow charts in Python. Everything looked great in local demos and leadership was thrilled but as soon as we tried pushing these agents into actual production for client-facing workflows, the real pain started and it had almost nothing to do with the LLMs themselves. The bottleneck shifted overnight from "how do we build this agent" to "how do we safely deploy, audit, and govern it." Suddenly, we were dealing with messy manual approvals, no clean way to roll back when a tool call broke, zero visibility into who owned which running agent and security teams panicking about identity management and raw cloud keys. We spent weeks trying to glue together custom GitHub Actions, Docker scans and ad-hoc logging scripts just to get basic deployment guardrails in place. It made me realize that the agent ecosystem is repeating the early DevOps transition. Building the code is the easy part now; managing the pipeline around it is where projects quietly stall out. Tools like Lyzr Control Plane, Microsoft's new reference architectures and dedicated agent gateways are starting to treat agent deployment like actual enterprise software with evaluation gates and automated pipelines which feels like a necessary shift. Are you building custom CI/CD wrappers internally to scan and deploy your agents, or are you moving toward specialized orchestration and control layers?