The AWS Developers Podcast

Coding agents write code faster than ever, and the bottleneck in shipping software has moved to validating it. A 15-minute build gives an agent 4 feedback cycles an hour. A 5-minute build gives it 12.
Romain sits down with Dan Manges, co-founder and CEO of RWX and former CTO of Braintree and Root, to talk about what CI/CD looks like when agents are the main users, and how the rest of the software development lifecycle has to change with it.
Key topics covered: - Why traditional CI/CD pipelines slow down coding agents - Graph-based execution and content-based caching, compared to Docker's lineage-based caching - A CLI-first workflow: run CI against local changes without committing or pushing, and get a green PR check right away if the agent already ran it - Sandboxes for the agent's inner loop, MCP, semantic outputs, and retrying only the failed tests - Building RWX on AWS: a distributed container runtime with no image build and push, a hot pool of EC2 M8a instances, single-tenant VMs, and self-hosted runners in customer VPCs - Rethinking the SDLC from first principles: give agents the context to get it right, throw code away and fix AGENTS.md instead of iterating - Software factories for internal tools versus human review for mission-critical code - Turning large changes into reviewable work by building the whole change first, then extracting shippable pieces - Guardrails for agent-driven changes: CloudWatch, gradual rollouts and feature flags - Results: pipelines about 2x faster on average, and about 35% lower cost - Is infrastructure as code still necessary?
Chapters: 00:00 Rethinking engineering for AI agents 00:52 Meet Dan Manges and his engineering journey 02:59 Why traditional CI/CD becomes a bottleneck 05:34 Faster feedback loops for coding agents 08:02 Graph execution and content-based caching 13:19 A CLI-first, agent-native CI workflow 15:33 Unifying agent development and CI environments 17:45 Building RWX on AWS 21:33 Rethinking the SDLC for agents 24:33 When humans should review agent code 27:07 Giving agents the context to get it right 29:29 Sharing agent skills and team practices 35:50 Building software factories with agents 40:32 Turning large changes into reviewable work 46:20 Guardrails for safe agent-driven changes 48:06 Measuring CI speed, cost, and agent benefits 51:10 Signals your CI pipeline needs rethinking 54:02 Is infrastructure as code still necessary? 56:38 Closing thoughts and recommendations

An agent harness is everything around the model that turns it into a working AI agent, and most agents are missing key pieces of it. In this episode we go through five of them: context management, sandboxing, fine-grained policies, fast decision models, and a harness you can run in production.
Romain sits down with Morgan Willis, Senior Developer Advocate at AWS, and Arron Bailiss, Principal Engineer and Tech Lead for Strands Agents, to look at what it takes to build a state-of-the-art agentic system today, and how Strands Agents grew from an agent loop SDK into a full toolkit for it.
The 5 pieces: 1. A harness with built-in tools, memory and a CLI, open source as Strands harness or managed as AgentCore harness 2. Context management that offloads tool results and compacts history, for 28% lower token cost on the same model 3. A sandbox that limits what files and network your agent can reach (Strands Box) 4. Policies that remember what the agent already did, so it can't issue a refund before checking eligibility (Dogwood) 5. Small decision models that pick between options in milliseconds instead of a full LLM call (Strands Decider)
Plus: real-time voice agents with bidirectional streaming, and why Morgan turned a Big Mouth Billy Bass into one.
Chapters: 00:00 Strands one year after going open source 02:45 Why model-driven agents get better with every model 13:50 What is an agent harness? Model plus harness 18:20 Piece 1: a harness with tools, memory and a CLI 25:50 Piece 2: context management and 28% lower token cost 29:45 Piece 3: sandboxing agents with Strands Box 31:00 Piece 4: Dogwood policies that track what the agent did 42:45 Piece 5: Strands Decider and small decision models 51:15 Bonus: voice agents with bidirectional streaming 56:45 What's next: agent safety and self-improvement

In-flight entertainment now runs on a cloud. Not the one on the ground, a cloud inside the aircraft. A 300-watt onboard data center running Kubernetes, managed databases and observability, built to survive vibration, fire and electromagnetic interference. Then AI agents on top so airlines can talk to their fleet.
Romain sits down with Frédéric Faria, Chief Technology and Engineering Officer, and Yann Nicolas, Chief Data and AI Officer, at Thales In-Flight Experience, to unpack FlytEDGE, the cloud-native in-flight entertainment platform.
Key topics covered: • Why most aircraft still get content by USB drive, and how connectivity changes that • Designing an aviation-grade onboard data center: about 100 TB of storage in a 300 W budget • Vanilla Kubernetes, Helm and GitOps for updates and rollbacks in flight • An open onboard SDK with REST APIs for flight data (position, speed, origin, destination) • Multi-tenancy, data minimization and forensic watermarking for studio content • Global delivery with Amazon CloudFront and caching once per aircraft • A conversational ops agent built with Amazon Bedrock AgentCore, AgentCore Gateway and Strands Agents • Rolling out Kiro to hundreds of engineers, beyond software teams
Chapters: 00:00 Intro: your in-flight entertainment, now in the cloud 00:51 Welcome: Thales In-Flight Experience, 2 million passengers a day 07:24 What changed: connectivity and the end of the USB drive 12:36 A data center on board: 300 W, 100 TB, Kubernetes 18:00 Vanilla Kubernetes, Helm and GitOps at 35,000 feet 27:07 Opening the platform: onboard SDK, aircraft APIs, partners 32:52 Multi-tenancy, privacy and global delivery with CloudFront 41:19 DRM and forensic watermarking 47:35 AI agents for airline ops with Bedrock AgentCore and Strands 1:00:42 Kiro across the engineering org, books and dev tools

Is AWS Lambda still serverless? With Managed Instances, MicroVMs, and Durable Functions, Lambda can now run workloads it was never designed for. AWS Serverless Hero Yan Cui explains what changed, what you give up, and why the core serverless promise still holds.
Romain sits down with Yan Cui, AWS Serverless Hero and independent consultant, to discuss how Lambda grew into general-purpose compute and where it fits for AI agents.
Key topics covered: • Why pay-per-use is no longer the whole story, and why that's fine • Lambda vs Fargate vs ECS: packaging a container image is not running a container • Lambda Managed Instances, the 90-minute timeout, and Durable Functions for jobs that run up to a year • Cold starts: why they were never as big a problem as the conversation around them • Lambda MicroVMs as a second sandbox for AI agents and CI/CD runners • AgentCore Runtime vs Lambda MicroVMs for agentic workloads • Observability for event-driven architectures, and what Lambda still gets wrong • Yan's rule: every box on your architecture diagram needs a reason to exist
Chapters: 00:00 Highlights 00:55 Meet Yan Cui, AWS Serverless Hero 05:23 Is Lambda still serverless? 14:36 Lambda vs Fargate vs ECS: where to draw the line 20:46 Lambda Managed Instances explained 24:10 Are cold starts finally dead? 29:45 Lambda MicroVMs, AI agents and S3 Files 37:51 AgentCore Runtime vs Lambda MicroVMs 45:28 What serverless still gets wrong, and observability 56:07 Yan's advice, book pick and AI security worries

Is Jev just a smarter if-statement? That's the question everyone's asking about TypeSafe AI's new "System One" model. In this episode, we dig into what Jev actually is: a model that makes fast, typed decisions your code can use directly — and why it points to a bigger shift: stop defaulting to one giant LLM for everything.
Romain sits down with Mike Chambers, Senior Developer Advocate for Generative AI at AWS, to discuss Jev, decision models, and how they complement the LLMs we already know.
Key topics covered: • What Jev is, and why it is not just another LLM • The three primitives: Noul, Choice, and Score • Putting gated decision points inside an agent with Strands • Fast model routing, and why cost and speed are the real win • Why the right tool for the right job is back
Chapters: 00:53 Welcome, and what is exciting in the Bay Area 04:13 What is Jev? 05:24 Why it is not just another model 09:53 A decision model: moving logic back into code 17:31 System One, and Thinking Fast and Slow 19:02 The three primitives: Noul, Choice, Score 22:29 Accessing Jev: API, OpenRouter, SDKs 26:22 Use cases: gated decision points in agents 35:19 Model routing, cost and speed 40:17 Can Jev drive a DeepRacer, and other models to consider