Hi guys :) New to the channel here. Writing from Argentina, sorry if my english is not that polished. I work at a small startup and have been diving into platform solutions for automating our ops stuff. So we looked at Backstage initially but honestly it felt like huge overkill. We basically just wanted deployment automation, so I ended up rolling my own solution - nothing fancy, just a UI for devs to plug in their service configs and a backend that handles manifest generation and deployments. But lately I've been thinking this solutions are going to be totally obsolete soon... I feel like most of the low level firefighting daily tasks can be scripted, wrapped in an MCP and exposed to an LLM. Has anyone here been experimenting with agentic platform engineering workflows?

I would not worry about things getting obsolete at the moment. Even in your example, someone has to build the wiring (MCP). The other thing is UI - does a chat provide the best experience when it comes to deploying things? Questionable. So good UX and clarity of the outcomes are still important in my opinion.

We're piloting an assistant that answers questions like: 1. What version of (something) is deployed to (environment) 2. Can I deploy the latest version of (something) to test 3. Can I create a new deployment project to deploy a (thing) to (tech stack) The crucial part of this is to apply the user's permissions to all these requests (we do it through the API, so this is baked in) and to confirm everything back to them before you do it. Like intent-based leadership... "Would you like me to deploy v1.2.3 of the Shopping Cart API to Staging?" That way folks can back up if they asked for something they realize they don't want, or if there's ambiguity (like Shopping Cart API, Shopping Cart UI).

I guess that what's already built and working won't need fixing or become obsolete. But my head immediatly goes over to this thought: if i had to develop a solution for the same problem at another company I'd probably go for an agentic design. If i could do it over, I'd just stick an MCP on some custom scripts, expose a schema and let the agent handle it. I don't have experience developing MCPs, so there may be a learning curve there.

We built an entirely SaaS-ified, multi-tenant backstage product (Flightdeck) with AI functionality baked in. While taking customer requirements for their AI and UX expectations it became clear Backstage was not a good fit for what most customers were looking to do. We came to the same conclusion you did, that the right approach in a post-AI world is an agentic model, leveraging many MCP servers (eg the tools already in your environment) with interfaces delivered inside of a users existing workflow. Lowers the toil for everyone involved and is vastly superior UX for most teams.