Kilo

A quick recap of last week's AI Engineer by Mistral. Great event Alisdair Broshar and crew 👏 Read more: https://lnkd.in/eFru4APm

Kilo now supports Sign in with ChatGPT. 🔑 You can create a Kilo account with your ChatGPT credentials, then connect your ChatGPT plan and use it across VS Code, JetBrains, the CLI, Cloud Agents, Code Reviews, and the mobile app. You can also add your ChatGPT plan as a BYOK provider, and your requests draw on the subscription you already pay for instead of per-token API billing. If you start a Cloud Agent from your phone, it runs on the same plan you use in your editor, with no extra OpenAI API key required. Your ChatGPT plan can sit alongside the 500+ other models in Kilo, and you can switch between them at any point in a task. For models your subscription doesn't cover, you pay the provider's own per-token price with nothing added on top. The full post covers how it works! (link in comments 👇)

Kilo showed up at two Builders & Brews events last Friday: Job Rietbergen spoke in Amsterdam, Brian Turcotte spoke in NYC. Great company on stage too: JetBrains, Agentic AI Foundation, Moyai, LangChain, Nebius, and Tavily. Builders & Brews is a running series from Nebius and Tavily across the world, and these two were part of the Nebius x NVIDIA online hackathon, with some great prizes on the line. Builders who join can use Nebius Token Factory inside Kilo. Time to build!

Another model shipped this week, and someone on your team probably wants to switch tools again. Emilie Schario wrote a short letter on the Kilo blog for engineers who feel like they can't keep up with the pace of AI launches. She argues that you can skip most of the launches and still get better at using AI, as long as you test new tools against work you already do. She also suggests a weekly time budget for experimenting and a clear rule for deciding when a tool has earned a place in your workflow. Check out the full post: https://blog.kilo.ai/

A security team spent hours on July 24 trying to exploit an old bug in OpenAI's help forum and couldn't get it to work. That same evening Anthropic shipped Claude Opus 5, and by the next afternoon the team had reached OpenAI's internal code repository. The bug, the researchers, and the target all stayed the same overnight, so the only variable was the model. The fix that would have prevented it had existed for over a year, and no standard scanner would have caught its absence, which says a lot about how most teams check their dependencies today. We wrote up what happened and why it changed how teams should think about locking into one model when the frontier shifts daily. Full post here: https://lnkd.in/egFACaf4

Meet Maisie Connaghan and myself at AGNTCon + MCPCon over the next two days. Learn all about how Kilo and Anaconda can accelerate your agentic development. And oh, we have LEGO! Join the raffle!

#AGNTCon + #MCPCon Europe is kicking off at RAI Amsterdam today, and we're proud to be among its sponsors. If you're around, make sure to stop by our booth #S8 to meet the Anaconda team in person and enter the raffle for a chance to win fun prizes! The event agenda is packed with conversations around what it really takes to build and scale agentic AI systems in production, from orchestration and infrastructure to evaluation, resilience, observability, and debugging. Hope to see you there 🌷 Schedule ➡️ https://bit.ly/4fIOhmu Register ➡️ https://bit.ly/4xRZN6j cc Job Rietbergen Maisie Connaghan

📱 Kilo App is live on Product Hunt! Start coding agents, check and manage remote sessions, or review PRs from iOS and Android. If you’d like to support us, an upvote or comment would mean a lot. Link in the first comment 👇️

Kilo for iOS and Android is live. Kick off cloud agents, review PRs, and check in on running tasks from your phone. Coding doesn't stop when you close your laptop. Support the launch: https://lnkd.in/g7zBvBN7

An AI agent told a user its bug was filed. The tracker still had nothing in it, and the automated check passed anyway, because the agent's reply happened to contain the word "created." Florian Hines built a tutorial around that problem. He wrote a small issue-triage agent, broke its tests on purpose, and then documented how to fix them so they can tell an agent that did the work from one that only said it did. Along the way he gets into questions most people don't ask until something goes wrong in production: how two conversations can end in the same state and only one of them deserves to pass, and how two prompt versions can score identically while one is quietly worse on the cases you care about. If you have shipped anything with an agent in it and you still decide whether it works by reading the output, his post gives you something concrete to check instead. He also packaged the lessons as a skill, so your coding agent can walk you through them if the full read is too long. Florian wrote and built all of it, and it's here: https://lnkd.in/gFjzTpgF