Texting Your Agent: TechCrunch's Survey of Assistants That Live in iMessage, WhatsApp and SMS Shows Where the Permission Problem Moves
An October 3 TechCrunch roundup lists close to twenty assistants you reach by text message instead of an app -- Poke, Instinct, Ollie, Wajo's Fo, Town and more. The interesting part is not the interface but what each one is allowed to touch: inboxes, calendars, phone numbers and payment cards.
GitHub Trending: Cloudflare Open-Sources Its Internal Agent Workspace, and Agent-Reach Gives Agents a Way Onto Social Platforms
Two repositories on GitHub's October 4 daily trending list show opposite answers to the same question -- how much should an agent be allowed to reach? Cloudflare OS starts agents with zero ambient permissions; Agent-Reach, the day's biggest gainer, wires them into Twitter, Reddit and YouTube via cookies and scrapers.
Framework Watch: Claude Code 2.1.289 Closes Five Ways Deny Rules Could Be Skipped
Released October 3, 2.1.289 is mostly plugin and mod stability fixes, but five items matter to anyone who relies on permission rules: env-var prefixes and bare assignments under sandbox auto-allow, symlinked IDE files, managed-machine compound commands, and a plugin that could rewrite an org MCP server's tool descriptions.
Trend Watch: OpenAI's Departed Safety-Reports Lead Explains Himself -- 'The Company's Culture Is Broken'
Days after his exit was reported, David Robinson, who led the safety reports behind OpenAI's major launches, published an October 3 essay in The Atlantic arguing that trial-and-error development 'guarantees periodic failures' and that frontier labs should run like nuclear plants or busy airports. OpenAI says it pauses training when it needs to.