Cloudflare pipes stack traces straight into coding agents

Detect and send production issues straight to your agent — closes the loop from monitoring to remediation: built-in error monitoring in Cloudflare Workers groups production failures, then ships stack traces, logs, traces and app context to a coding agent that investigates and opens a pull request.

The interesting part isn't the agent, it's the grouping and context assembly — the unglamorous work that usually determines whether an automated fix is plausible or noise. Review discipline is now the bottleneck, not triage. — blog.cloudflare.com

Introducing cf: the agentic CLI for the entire Cloudflare API — mirrors the entire Cloudflare API in one command-line tool, with programmatic TypeScript configuration and an open-source SDK generator called Forge.

Worth a look even if you don't use Cloudflare: generating full API surface coverage is the current prerequisite for agents that can actually operate infrastructure rather than guess at it. — blog.cloudflare.com

RUMOUR

Microsoft abandons personal AI chatbot race with Copilot reboot — reportedly retreats from the consumer assistant fight and resets Copilot. For anyone building on Copilot integrations or betting on its API surface, a reposition is a roadmap risk — and it leaves ChatGPT and Gemini with the personal-assistant field to themselves. — sbulaev

Once Claude can measure something, it can make it faster — walks through how instrumentation, not model swaps, drove Claude's latency wins. The useful read is which bottlenecks and metrics actually moved the needle — a reminder that most "AI is slow" complaints are ordinary systems problems wearing a new hat. — matthieu_bl

The Pentagon wants $30 million to build an AI-powered lie detector — requests $30.3 million over five years for Polygraph+, a next-generation program built on ML scoring algorithms and "standoff sensing" — measuring subjects at a distance.

Before anyone treats a biometric score as evidence: ask about validation methodology, false-positive rates, and adversarial robustness. Polygraphs were already contested science; adding a model doesn't fix the base rate problem. — Amit Katwala

Who we become when we talk to machines — is Sherry Turkle's new book, which offers "a withering critique of chatbots and the antisocial dynamics she believes they encourage." Read it as product research if you ship conversational interfaces: dependency, social substitution, and the limits of simulated empathy are design decisions, not accidents. — Peter Dizikes | MIT News

Last Week in AI #345 - 5 new models, 9 misalignment incidents, some Dots — counts five new models, nine misalignment incidents disclosed by OpenAI, and Anthropic and OpenAI racing each other to smarter-and-cheaper. A fast scan rather than a deep read — the incident list is the part worth clicking through to. — Last Week in AI