The AWS Developers Podcast

Prime Video serves over 300 million monthly customers across 240+ territories and 8,000+ device types. Lilia Abaibourova leads the AI-native transformation there. Prime Video started early, adopting CodeWhisperer as soon as it became available in preview, progressing through Amazon Q Developer, and now using Kiro to put agentic development in the hands of engineers, product managers, and designers. Romain sits down with Lilia to trace that journey — from CodeWhisperer preview to Amazon Q Developer and now Kiro — and explore MCP as the organizational inflection point: teams run two-person pods with full decision autonomy, while managers oversee fleets of agents instead of just headcount. Key takeaways: • Prime Video adopted AI coding tools from day one — CodeWhisperer in preview, then Amazon Q Developer, and now Kiro — building institutional muscle even when specific features were later sunset • MCP was the inflection point for organizational (not just individual) transformation — connecting disciplines end-to-end and eliminating handoff bottlenecks • Code generation is now effectively solved; the new frontier is verification, validation, and code review at scale • Team structures are being actively experimented with — from 2-person pods with full autonomy to larger teams where managers leverage agent fleets • The biggest anti-pattern: waiting on the sidelines for a standard instead of learning through experimentation

I keep hearing 'we're building an AI platform' from enterprises. Most are stitching together demos. Thomson Reuters has been doing AI for 30 years. I sat down with Laura Skylaki (VP AI Engineering, Thomson Reuters) at AWS Summit London — and the gap between what they're doing vs. what most companies call 'AI maturity' is striking. In this episode, Laura breaks down how Thomson Reuters built an enterprise AI platform that serves 30,000 employees across three user personas — from business users to AI creators to developers — with 16 internal services, a $200M/year AI commitment, and a healthy skepticism about when agentic is actually the right answer. Key takeaways: • Evaluation before building — 'It's increasingly harder to retrofit an evaluation framework to an agent capability you already deployed.' They build evaluation benchmarks first to force understanding of the business problem, then build the agent. LLM-as-a-judge with human-in-the-loop calibration — and the judges themselves need continuous regression checks. • Know when NOT to use agents — 'Where we can go and write deterministic code with a logical rule-based approach, I don't think it makes sense to rely on the agent.' Not everything needs to be agentic. When tools exceed a dozen or overlap semantically, specialize into multiple agents instead of overloading one. • Engineers haven't written code manually in 3-4 months — Not experimenting; their AI coding tooling is that far ahead. The shift is toward system design, architecture, and clear structuring for AI assistance. • AWS under the hood with AgentCore and Strands — The platform runs on AWS with API Gateway, Lambda, Step Functions, and Amazon Bedrock AgentCore for agent deployments (runtime, memory, MCP marketplace, code interpreter, browser tools, policy and identity governance). They use Strands Agents SDK to build agents. AgentCore removes the infrastructure burden of long-running agent sessions. • Rethink workflows, don't retrofit — Teams reimagining legacy workflows from scratch for AI rather than bolting generative AI onto systems not built for it. 'If I was building this now for the first time, how would I build it?' The companies that succeed with AI are not fitting it into existing processes — they're redesigning the process through AI.

Brenn reached out via LinkedIn to share his experience with AI-driven software development and his new book, The Delivery Gap. Romain read a copy during a business trip and found it deeply aligned with how he guides customers through their AI transformation — and discovered a few new angles worth exploring, including a convergence on cost tracking that maps directly to Amazon's internal cost to serve software metric. In this episode, Brenn — Senior Manager at Delivery Hero (one of the world's largest food delivery companies, operating in 65 countries) — breaks down why most companies fail to see returns from AI coding tools despite individual developers feeling more productive. The core insight: generating code 10x faster means nothing if your verification infrastructure can't keep up. You're just driving 10x faster into a wall. Key takeaways: • The 96% waste problem — If you generate 100 PRs and only 4 make it to production and stay there, the other 96 are waste. Measuring PRs created is meaningless; measure what ships and survives. • The verification triangle — Your delivery speed is governed by verification infrastructure, not generation speed. Banks can't release faster than they can audit. Find your constraint — that's where investment should go, not more coding tools. • Cost per accepted change — Total token costs + human time for all PRs, divided by changes that reach production and stay there. This single metric reveals where waste accumulates and aligns with Amazon's cost to serve software model. • Specs as alignment documents, not source code — Specs align humans and AI on intent and why, not for deterministic code generation. The same spec produces different software each time. Focus on why; let the AI document the what. • Keep agents small and focused — Every MCP server re-injected into context is a cost multiplier per turn. The smallest, tightest, most precisely aimed agent outperforms a Swiss Army knife agent on both cost and accuracy. Apply cost per accepted action to measure agentic ROI.

Most MCP tools today return text. But what if your agent could render a chart, a form, or a full dashboard right inside the chat? That's what MCP apps do — and they're already live on ChatGPT, Claude, and Amazon Quick. Romain sits down with Luigi Pederzani, co-founder of Manufact (the company behind mcp-use, 10K+ GitHub stars), to explore MCP apps — the first official extension of the MCP protocol that lets servers return interactive UI inside AI chat interfaces. Key takeaways: • What MCP apps are — Standard MCP returns text and actions; MCP apps let a server send back a piece of interface (a form, a chart, a dashboard) that renders right inside the chat. It became the first official extension of the protocol this year, growing out of MCP-UI. • UI drives retention — Amplitude saw 2x retention for users exposed to a chart-rendering MCP app versus text-only responses. UI lets software products stay experiences, not just systems of record. • AI apps are the new browsers — Extending Paul Graham's thesis, every software product will be rendered inside AI chats, pulling data and structure from different sources the way we switch tabs today. • Building with mcp-use — Reuse existing React components with minimal changes; the useMcp hook bridges tool arguments (filled by the agent) into component props. The server stays a normal MCP server, the client is the host, and the view runs in a sandboxed iframe. • Interactivity and safety — Iframes are battle-tested and the model-to-server communication is standardized. UI can send events back to the model, so clicking a chart element can trigger another tool call in the chat. OAuth is now standard for production MCP servers. • Tool design best practices — Don't wrap OpenAPI specs directly as MCP. APIs are granular and atomic; MCP tools should serve a task end-to-end so agents don't get confused on ordering. Limit the number of tools exposed; progressive disclosure is now handled by the major clients. • MCP as the A2A protocol — A2A never really landed, and MCP is becoming the agent-to-agent protocol, with companies embedding an agent as a single tool whose main argument is a prompt. • Getting to production — Start with a plain MCP server, then add UI. Skills are now part of the product, and Manufact focuses on agent-readiness of the SDK, testing across clients, deployment, auth, observability (OTEL), and per-tool analytics. • AWS integration — mcp-use can sit on the server side while AgentCore Gateway sits in front to handle enterprise concerns like auth policies and routing. • What's next — Exposing skills directly from MCP servers (rather than decoupled files), and the next stateless-by-default release of the protocol.

John Sexton and Aaron Tummon from Genesys join the show to share hard-won lessons from building and operating Cloud Copilot — an agentic AI layer serving 2 million users across 21+ AWS regions. Genesys powers customer experience for brands like Virgin Atlantic, Vodafone, and HSBC, and their copilot helps admins, supervisors, and agents work more efficiently through natural language. We cover the migration from Bedrock Inline Agents to Strands Agents, multi-agent orchestration with agents-as-tools, context management strategies, cost optimization, and the testing discipline required to keep agentic systems stable at scale. The 5 lessons: 1. Pick a framework that scales with you — Bedrock Inline Agents worked for 12–15 tools but became exponentially flakier beyond that. Strands Agents gave sensible defaults and room to grow without pinch points. 2. Separate orchestration from domain logic — Agents-as-tools creates a clean line between the orchestrator and sub-agents. You can pull functionality in and out per persona without destabilizing the system, and domain teams own their sub-agents independently. 3. Manage context aggressively — Long context windows for the orchestrator, stateless sub-agents, summarizing and sliding-window conversation managers, and strict control over what tools return. Every extra token in context degrades quality and increases cost. 4. Make prompt caching non-negotiable — System prompts, tool definitions, and conversation history rarely change between invocations. Enabling prompt caching delivered significant cost reductions with almost no effort. 5. Test relentlessly because prompt drift is invisible — One prompt change is never a breaking change; five accumulated changes are. A dedicated weekly Sentinel role investigates failures, and full test suites run on every single change.