OpenAI's software factory and Steve Yegge's Wheelhouse disagree about human code review, plus Amazon Quick agents that keep running after you close the laptop and an AWS pattern that turns IAM findings into pull requests.

A swarm of AI agents used a 25-year-old German wiki as a backchannel to share answers to the tests they were being graded on, plus GPT-6 Astra lands on Bedrock and Amazon Linux 2027 turns SELinux on by default.

Agents followed install commands in llms.txt files pointing at packages nobody owned, OpenAI cuts Cursor off in November, and GLM 5.3 Flash does the job for nine cents.

Opus 5 quietly stopped loading Brett's skills for a day, EC2 turns 20, AWS buys DuckLabs, and OpenAI's agents built themselves a message board.

Opus 5 has started taking actions nobody asked for, GLM-5.3 learned to plan complete exploit chains, and Claude Code is now 44 percent of the agent traffic hitting Hugging Face.

Kiro Crew imports your whole agent workflow on first launch, DynamoDB gets native vector search, and GLM's low reasoning setting turns out not to be low at all.

AWS Distinguished Developer Advocate Darko Mesaros on 1974 minicomputers, writing Rust with LLMs, and why solving code generation did not solve software engineering.

AWS Cost Explorer shows users a $2.5 trillion bill, Kimi K3 brings open weights to the frontier, and an OpenAI model escapes containment to hack Hugging Face.

Talking Cloud episode 52: OAuth for the AWS MCP servers, GuardDuty AI Protection, Loom for AWS, Bun's $165K Zig-to-Rust port, and OpenAI's GPT-Red.

Talking Cloud episode 51: Lambda microVMs, AWS Blocks, a pile of service deprecations, a silent EventBridge retry loop that 6x'd a Bedrock bill, and Midjourney pivoting to full-body ultrasound.