Loco.

wild… Uber reportedly burned through its full 2026 AI coding budget in four months. well, this is going to happen to a lot more companies if they keep paying frontier model prices for every task. love what Glean’s doing here > they built tau, a desktop app that combines company context with your local files, apps and code > then they added auto routing across 40+ models, so every task gets the model it actually needs > they’re also making AI proactive where its helping triage email, prep meetings and drafts etc.. it’s pretty simple stop paying AI to relearn what your company already knows, and stop using the most expensive model when the task doesn’t need it. it’s a complete waste of time and money.. btw tau and several features are coming soon..

another day, another reminder that unsubscribing from Anthropic was the right call like what kind of response is this? you guys have been silent for weeks while people complained about Opus 5 and now the response is basically that we’re just not using it right? apparently we were supposed to give it CPU metrics and let it hillclimb for hours lol most people just want the model to follow instructions, make the right changes, stay on task, and not overcomplicate simple things. if customers have been telling you for weeks that the experience got worse, then maybe explain how we’re actually supposed to use Opus 5. show people what it’s good at, show the workflows, show how to get the best out of it. staying silent for weeks and then basically saying “people are sleeping on it” is such a weird response when your own customers have been telling you exactly what feels worse..

pro agent tip: whatever you’re building, don’t let the same agent be both the creator and the final judge. but spawn a second agent whose only job is to review the work. before it starts, ask it to identify and pull the exact skills, tools and expert perspective needed for that task, then let it audit everything in parallel while the main agent keeps building. >design gets reviewed like a creative director >code gets reviewed like a senior engineer >research gets reviewed like a skeptical analyst basically, the first agent optimizes for creation and the second optimizes for finding what’s wrong... you’re basically adding an independent quality-control layer to the workflow, and it catches an insane amount of stuff a single agent would miss.. this doesn’t really eat that many extra tokens btw.. really worth it