This is part of a series of blog posts that stemmed from my first experiences with “vibe coding”. For more on what I did and where you can find the rest of the posts, please refer to the first post.
The most important takeaway for me for all of this is that Agentic AI (as it is in many places) is the answer to how the LLMs can be applied most efficiently. When I say agentic AI, I am referring to two specific things that will have an outsized impact on how AI can be used to develop software.
- Access to tools via MCP servers – MCP (Model Context Protocol) servers are essentially interfaces between AI Agents and API layers from platforms. Many of these that would be useful in Application development are already appearing (earlier this month AWS actually announced the release of an MCP server!)
- Memory – Memory means lots of things in agentic development from session memory to storing artifacts. What I think will matter in software development in the short-term is the ability for developers to modify the agents based on watching them run. For example, I learned the hard way that there are a few gaps in how the current LLMs go through an Angular upgrade (
Together these address two out of the three limitations that I raised concerns about in this series:
- Intern-like behavior (my blog post on the problem is here). With the ability to constantly tune agents based on what we learn we can correct for this though. If I was going to be migrating another 50 sites from Angular 6 to Angular 18 (the kind of thing that’s regularly part of our projects at Kyndryl), I would have written an agent that was aware of all the things it needed to check (like the size of icons in the mobile view).
- The fact that AI Assisted Development is limited to what it can do from the IDE (my blog post on the problem is here) is mitigated through agents that leverage the MCP servers above. I could create an agent that sat on top of the Claude LLM that could work with me on the infrastructure by using the MCP server above.
This blog post is the last one in my series of vibe coding reflections, but it’s the most important for me personally. As Kyndryl’s US Leader for App Development and App Modernization, I get to set a lot of how we prepare ourselves. We are already amassing the agents that our team has been using on projects over the last six months. Centralizing them and making them accessible to everyone. We’ve even invested in building some of the ones that are key to our software development projects. I can’t wait to be a catalyst for our customer’s development and modernization efforts.