AWS Agents Hit GA and Claude Code Gets Leaked

Top-Line Summary

AWS dropped three agent GAs in one week — AgentCore Evaluations, DevOps Agent, and Security Agent — all with multicloud and on-prem support, making it clear AWS wants to own the operational layer regardless of where your workloads run. The other headline: Anthropic accidentally leaked the entire Claude Code source, and what people found inside was surprisingly pedestrian — regex-based sentiment detection and anti-distillation poison pills.

Show Video

The “Pre-Show” Context

Brett and Travers were 15 minutes late because they were watching the Artemis II launch — the first crewed lunar mission since Apollo 17 in 1972. Four astronauts including Christina Koch and Canadian Jeremy Hansen headed for a 10-day moon flyby. Weather was 90% go, and the last surviving moonwalker made a cameo on the broadcast.

The Engineering Rundown

• Bedrock AgentCore Evaluations (GA) (02:00)

  AWS now has built-in quality gates for AI agents — online evaluation samples live traces in production, and on-demand evaluation plugs into CI/CD. Thirteen built-in evaluators cover task completion, safety, and tool usage, with custom evaluators via Lambda for domain-specific scoring. Brett admitted he keeps saying he’ll dig into AgentCore and never does, while Travers noted AWS is pushing news items for it practically every week. The unspoken message: if you’re shipping agents without evals, you’re shipping hope.

• AWS DevOps Agent (GA) (04:00)

  AWS’s AI SRE play went GA after a free preview since re:Invent. The big addition: multicloud and on-prem support, so it’s not locked to AWS-only environments. Travers, doing SRE work currently, saw the immediate appeal — kick off agents to pull logs and triage across your stack. Pricing landed at $0.0083 per second across all task types. Region availability expanded to six AWS regions but still no Canada, which prompted Brett to consider deploying a personal project in us-east-1 just to test it. PagerDuty, Grafana, EventBridge, and Azure DevOps integrations sealed the deal — both hosts committed to actually trying it this time.

• AWS Security Agent — On-Demand Pen Testing (GA) (11:00)

  Autonomous pen testing across AWS, Azure, GCP, and on-prem, priced at $50 per task hour with a two-month free trial. Travers dug up the pricing estimates: a small web app runs about 8 hours ($400), a medium e-commerce platform 24 hours ($1,200), and a large enterprise app 48 hours ($2,400). The real question is whether this replaces your pen testing vendor or augments them — Brett’s take is it handles the commodity scanning work, freeing your security team for context-heavy attack scenarios. For smaller orgs that can’t justify annual pen test engagements, continuous testing versus point-in-time snapshots is a fundamentally different security posture.

• Console Visibility Customizations (17:30)

  Brett’s favorite announcement of the week, and it’s not even AI-related. You can now hide AWS services and regions from the console per-account via CloudFormation and CDK. Brett already had a customer in mind — one where he regularly fields support tickets from people alarmed by red permission errors on services they were never meant to access. Travers estimated it would kill a hundred messages a week. Brett joked he needs an FAQ bot that just replies “it’s okay that it’s red, you don’t have access” so he can move on with his life.

• Personal Projects & Tools (21:30)

  Both hosts spent time talking shop on their personal software builds. Brett turned off two SaaS tools this week — Jira and a newsletter platform ($35/month, replaced by SES at 58 cents/month). He’s shipping his first non-SaaS newsletter tomorrow morning and eyeing the podcast hosting platform next ($20/month, replaceable by S3 + CloudFront). Travers has been reading Vibe Coding by Gene Kim and Steve Yegge and building his own OpenClaw-style clone backed by a PARA knowledge system with vector search. A former colleague of Brett’s published a game on Steam — a top-down 2D shooter — and Brett’s trying to get him on the show. Both agreed: the era of personalized software is here, regardless of your technical skill level.

• MIT — AI Warehouse Robot Traffic System (40:00)

  MIT and Symbotic built a hybrid system — deep reinforcement learning for prioritization, classical planning for movement — that achieved 25% throughput gains coordinating warehouse robots. The winning pattern was AI augmenting proven algorithms, not replacing them. Travers noted they included a random search component for chaos resilience. Brett wondered whether the Petri dish results would hold in a real warehouse with a thousand robots — likely yes, but with messier edge cases.

• Ethan Mollick — “Claude Dispatch and the Power of Interfaces” (43:00)

  Mollick’s argument: AI’s limitations are mostly an interface problem, not a model problem. The Claude Dispatch example — texting Claude from your phone while it controls your desktop — prompted both hosts to reflect on their own UI preferences. Brett tried Dispatch a few times but found the asynchronous nature jarring (“I texted it and nothing happened — are you doing anything?”). Travers pointed out you need some kind of interface agent for progress tracking without constant notifications. Brett’s slightly cynical take: Dispatch and the new Loop feature feel like Anthropic’s response to the OpenClaw hype. They coined “dim workshop” as the Level 4.5 between hands-on and dark factory.

• Mario Zechner — “Thoughts on Slowing Down” (52:00)

  The anti-hype counterweight. Zechner argues AI agents ship code faster than teams can maintain quality, citing Microsoft’s 30% AI-written code alongside declining Windows quality. Brett’s response: Windows sucked when he was installing it from 47 floppy disks, so blaming AI is a stretch. His hotter take: setting daily limits on agent-generated code is swimming upstream — you’d better figure out how to operate in this space or get left behind. Travers framed it as a review and testing problem, not a generation problem. Both agreed the human is still responsible for what gets shipped, agent-written or not. The conversation pivoted into how AI has actually improved merge request quality — agents write better commit messages than most humans who just type “fixed bug.”

• Google Research — TurboQuant (59:00)

  Google figured out how to compress AI model key-value caches down to 3 bits with zero accuracy loss — 6x memory reduction and 8x speedup on H100 GPUs. Travers noted this caused a stir in memory stock markets, but his counter-argument was simple: why wouldn’t they just use 6x more RAM instead of 6x less? The real impact is running inference on smaller devices, which changes the economics of edge AI deployment.

• AWS Architecture Blog — Architecting for Agentic AI (01:01:00)

  AWS’s reference architecture for agentic AI development. The diagram shows Kiro submitting code to local Lambda and DynamoDB, iterating locally, then pushing through CDK via CodeBuild and CodePipeline. Brett’s take: this is CI/CD 101 — fail fast, fail close to the developer — dressed up in agent terminology. He then roasted CodeBuild for being painfully slow, recounting a recent class where students asked “is it working?” and he told them to go get coffee. Travers agreed: CloudFormation plus CodeBuild might be the slowest deployment loop imaginable.

• Claude Code Source Leaked (01:05:30)

  Anthropic accidentally leaked the entire Claude Code source via a deployment file on Cloudflare R2. Travers walked through the findings: sentiment analysis is just regex matching against a list of angry words, and there are anti-distillation measures that poison output with fake tool calls when they detect someone training against the model. Repos porting the code to Rust and other languages appeared almost immediately. Brett’s cynical take: Anthropic blamed “human error,” but he wonders if an agent made the mistake — which at Anthropic might be the same thing.

Off-the-Clock Recommendations

• Project Hail Mary (2025 film) — Brett’s pick. Fantastic adaptation of the book, Ryan Gosling likely award-worthy. Brett paid for D-BOX seats that didn’t work and then couldn’t reach a human through the theater’s chatbot. The irony was not lost.
• The Mummy Returns (2001) — Travers’ anti-recommendation. The first Brendan Fraser one is fun; the sequel is not.
• Super Mario Bros. Movie: Mario Galaxy — Travers was heading to see it right after recording. Brett passed.
• Vibe Coding by Gene Kim and Steve Yegge — Both hosts reading it. The feedback loop framework resonated with their personal project workflows.