What's New In Data

The episode contains solid technical insights about the model-driven approach to agents, tool-use patterns, and the shift from RAG to retrieval-as-a-tool, but much of the conversation involves context-setting, product explanation, and general discussion about the AI landscape rather than densely packed novel ideas. The section on vibe-checking and evaluation frameworks adds practical value, but extended portions are devoted to comfortable affirmation rather than challenging new claims.

“the model driven approach, because we were seeing that actually, we didn't need a lot of complexity”

“we're seeing a lot more of what I call retrieval as a tool as opposed to retrieval, augmented generation”

While the framing of retrieval-as-a-tool over RAG is reasonably fresh, and the multi-agent organizational perspective (two-pizza teams) adds some originality, much of the core thesis - that models are now powerful enough for tool use, that MCP enables standardization, that evals are hard - reflects emerging industry consensus rather than contrarian thinking. The naming bot anecdote is creative but feels somewhat tangential. The episode doesn't deeply challenge prevailing assumptions.

“don't actively give information to the model that might be too much information or even information that it doesn't need”

“multi-agent is sometimes more of an organizational solution where you have five different teams that want to own some experience”

Clare Liguori is a Senior Principal Engineer at AWS with 11 years at the company and direct responsibility for Strands Agents, Q Developer, and agent-related work. She has shipped real systems at scale and speaks from deep operational experience. However, she is primarily a product/engineering leader rather than a founder or business operator, which slightly limits the perspective to infrastructure/tooling rather than end-to-end business outcomes.

“I've been at AWS for almost 11 years now”

“When I started building AI agents back in early 2023”

The episode includes some concrete details (Strands launched in May, Claude 3.7 training data cutoff in March, agents modeling 6000+ AWS API tools, Aurora SQL example, RDAP domain lookup discovery) but relies heavily on abstraction and framework discussion. Many claims about MCP adoption, the agent ecosystem, and testing challenges are stated without supporting metrics, customer examples, or failure cases. The naming-bot walkthrough is specific but brief.

“we have an agent internally that models, more than 6000 tools”

“their training data ended in, in March of this year. Strands came out in May”

The host asks generally solid setup questions and occasionally pushes into specifics (evals, testing, tool selection), but rarely challenges Clare's framing or digs into tension. The conversation is warm and affirming but lacks the edge of genuine disagreement or sharp follow-ups that would expose assumptions. For example, when Clare mentions agents as the new libraries, there's no probe into the sharing/versioning/trust problems inherent in that model. The interview feels more like a guided tour than an interrogation.

“Well, a couple are one, it's not tied to any the particular model”

“And that's a great call out that, these lines are now being trained on using tools”