I got put into a meeting with a networking team member and three people from AWS, talking through our current AI use and where we want to take it. Nobody assigned me to lead it going in. About ten minutes into the technical parts, the networking guy turned to me and asked me to take it from there.
Research, fast
I hadn't worked with AgentCore, Bedrock, or Lambda before, so the days before the meeting went into figuring out what each service actually does and how it maps onto the MCP work I'd already built. The immediate problem they solve: my MCP servers are currently limited by not having a proper LLM API endpoint of their own. AgentCore gives us that, and it stays inside our existing AWS budget instead of becoming a separate line item somebody has to justify. It's also the missing piece for the PDF-import problem — running a PDF directly through a foundation model sidesteps the extraction ceiling that every coded approach ran into, and AgentCore is what would actually let us run that.
Weighing Lambda
I also looked at hosting the MCP servers on Lambda specifically — the tradeoffs around token efficiency and cost efficiency compared to how things run now. Lambda turned out to be the answer: it's how the servers would actually run and stay accessible to the whole team, instead of something only on my laptop.
Explaining, not demoing
I didn't run a live demo in that room — what I gave was a high-level walkthrough of all three MCP servers I've built so far: the PDF-import pipeline, ready to go once we have AWS access, and the two dev-focused tools — one that answers questions about our database and schema, one that uses SVN to answer questions about recent changes to files. All of it in natural language. Hosting them on AWS means they stop being something only I can run locally and become something the rest of the team can actually use: asking about a complex schema relationship that isn't obvious from reading the code or querying SSMS, or asking whether a recent commit might have introduced a regression.
The actual demo came later
The real demo happened separately, at our internal AI meeting, presenting to a senior dev. He liked it — and framed the value in a way I hadn't quite put into words myself: that this is a step in the right direction for proper system documentation. That lines up with the design system work I've also started leading; both are, underneath it, about turning tribal knowledge — schema relationships, commit history, component patterns — into something documented and shared instead of something that only lives in my head.
What this says about where this is going
AWS ended up resolving more than one open problem in a single meeting — the missing LLM endpoint for the PDF pipeline, the hosting story for the dev-workflow servers. It felt like more than a process improvement; it felt like a real step toward launching our first actual AI product.
Nobody planned for me to lead that meeting. It happened because the results from the last few months made it the obvious call. That's become the actual pattern here: the more the AI work holds up in practice, the more I get handed the next piece of it to lead.