BuildersVercel fenced its agent sandboxes with a network wallA box that isolates the processor but leaves the network open lets an agent carry the data out without ever leaving the box. Vercel adds a host-level firewall and keeps secrets outside the sandbox entirely.BuildersJetBrains put agents on a shared protocol and called it the LSP momentACP separates the editor from the agent: you swap agents without swapping environments and without losing project context. GitHub Copilot works as a protocol server, Claude Code connects through an adapter.BuildersCloudflare shipped a browser just for agents - cheaper, but slowerKitesurf isn't Chromium. It's built from open parts of Firefox and Servo, and runs straight on the workers of Cloudflare. By their measurements it eats up to seven times less memory - the price is a response about twice as slow.BuildersThe protocol for AI tools threw out the session - and that cuts the cost of running your own serverThe new Model Context Protocol specification drops the handshake and the session ID - every request now carries everything it needs on its own. Cloudflare shipped its implementation on 6 August. Three of the old capabilities enter a 12-month corridor before removal.BuildersGitHub slices the giant AI pull request into layers: gh-stack turns 1700 lines into a reviewable ladderAgents write code faster than a person can review it - and it all lands in one giant pull request. GitHub released gh-stack: a command-line extension that splits the work into ordered, independently reviewable layers. A change in the middle propagates up the stack on its own.BuildersCloudflare gives the agent a computer: files, git and shell without a container for everyoneAgents Week opened with empty talk. @cloudflare/computer is the first real meat - an open library that gives the agent a file system, git, and code execution. The lightweight mode carries most of the work; a real Linux container only fires up for the heavy stuff.BuildersThe harness beats the model: GitHub on where AI productivity actually hidesAn essay on the GitHub blog from 27 July says something simple and rarely said out loud: the biggest jump in working with AI agents comes from how well you master the shell around the model - the harness. The model is the smaller part of the equation. A thesis that matches the everyday experience of anyone building with agents.BuildersGitHub flipped the reflex: Dependabot now waits three days before suggesting a new packageFrom 23 July, Dependabot stops suggesting brand-new package versions right away - by default it waits three days. The rule only applies to routine updates; patches for vulnerabilities still arrive immediately. The reason is an uncomfortable truth about the supply chain: malicious versions live for hours before anyone catches them.
64 stories