Kilo

On Wednesday, we announced that Kilo has been acquired by Anaconda. This is a big moment for us. Kilo started as an open-source coding agent, and in a short amount of time became a #1 project on OpenRouter, processed 40T tokens, and brought in over 3M users. We built this with the community, in the open, and that hasn't changed. Joining Anaconda means we get to keep building the all-in-one agentic engineering platform we set out to make, but now with the resources and reach to bring it to a lot more developers and teams. Today at 1pm ET, our CEO Scott Breitenother and Anaconda's CEO David DeSanto are going live to talk about what this means for Kilo, for Anaconda, and for the future of agentic development. Sign up here: https://lnkd.in/eCUumKaX

This weekend we find out: can AI models predict football better than humans? For the entire World Cup, six open models — Kimi K2.6, MiniMax M3, Nemotron 3 Ultra, DeepSeek V4 Pro, GLM 5.2, and Qwen3.6-Plus — have predicted every match. Every model gets identical inputs: fixture and venue context, recent team form, season records, key players' club-season performance, and current match news. Never betting odds. Picks lock before kickoff. Two matches remain. The panel backs Spain (4/6) and France (5/6). The betting market is far less certain. The missing data point is the human side. Cast your picks before kickoff and see how the crowd stacks up against the panel. Link in the first comment 👇

We’re excited to announce Anaconda has acquired Kilo, the open-source, model-agnostic agentic engineering platform used by more than 3M developers! 🎉 This acquisition further cements Anaconda as the leading AI-native development platform. Anaconda has long-provided the trusted foundation from development to production. Now with Kilo, Anaconda provides a secure platform where agents can get work done at enterprise scale. A singular, enterprise AI platform, enabling security and governance – without the need to trade off velocity. Learn more in our press release → https://bit.ly/4w3yuEH 🐍

Big news: Kilo has been acquired by Anaconda, Inc.! 💚 We’re so proud of what our team has built - growing our open-source, model-agnostic agentic engineering platform from zero to over 3 million developers in only sixteen months. Becoming part of Anaconda’s trusted foundation is a natural fit with our open-source roots and mission to cover the full AI-native development lifecycle. This major milestone is only possible because of our customers, partners, and entire team. Our CEO and co-founder Scott Breitenother tag-teamed this blog with Anaconda CEO David DeSanto to share more about what this deal means. Learn more: https://bit.ly/4fkY7cV

Auto-approve lets your AI agent run commands without asking first. Kilo Sandbox makes sure that even if it tries something bad, your OS blocks it before it can touch anything outside the project. Speed without the risk. Learn more below 👇

Scott Breitenother's latest piece for Forbes Technology Council is live, and it's a good one for anyone trying to catch hidden pitfalls in AI coding right now. The core idea: writing code got faster, but debugging didn't get easier, it just moved. Agents nail the common case and struggle with the edge cases, so the time you save writing gets spent later on root-cause investigations where nobody on the team has full context on code they didn't actually write. Worth a read if you're trying to figure out where the "AI productivity gains" actually go once you look past PR count and cycle time. Link: https://lnkd.in/e_gFBn8G

We had Claude Fable 5 and GPT-5.6 Sol design the same 3 backend systems (webhooks, billing, permissions). Same prompt, no follow-ups, just a plan.md. GPT-5.6 Sol scored higher, 9.10 to 8.04. The reason why matters more than the score. GPT-5.6 Sol wrote a full plan, then launched sub-agents to audit its own work and revised based on what they found. Nobody asked it to. That's what caught the real bugs, like a billing dedup key that let retried events get charged twice. Fable 5 wasn't outclassed everywhere. Its plans were leaner and more buildable, and it finished in a third of the time at lower cost. The real takeaway: the audit step is what separates a plan that looks right from one that actually is. Write the plan, then have it (or a different model) review and revise before anyone touches code. That spend is nothing next to unwinding a bug like that in production. Full breakdown in the comments!

There's a gap in how most AI agent sandboxes handle credentials. Containers get isolated, torn down, spun back up, but the actual GitHub or Kilo token sitting inside one can usually do more than the agent needs, and if something goes wrong inside that box, that token is exposed. Evgeny Shurakov rebuilt how Cloud Agents handles this. The sandbox never gets your real token. It gets a short-lived one scoped to that specific container, and every outbound request gets checked before we swap in the real credential: did we issue it, is it going somewhere allowed, is it coming from the container it was issued to? If a token leaks, it only works from that one sandbox, for as long as that sandbox exists. Nowhere else, no other timeframe. Live now in Cloud Agents: app.kilo.ai/cloud

OpenAI's latest launch is a quiet admission that one "super-model" is not the future. Their new voice model, GPT-Live, doesn't actually answer your hard questions. When you ask it something that needs real reasoning or a web search, it hands the task off to GPT-5.5 running in the background and keeps the conversation going while it waits for the result. So they split the thing in two. One model is tuned for fast, natural, full-duplex conversation. A separate, heavier model does the reasoning and search. That's routing between specialized models, right inside their own flagship product. The most telling part is buried in the announcement: they designed GPT-Live so the reasoning model underneath can be swapped out for whatever frontier model is best at the time, while the interaction layer stays put. You don't build a swap-in system like that if you believe one model is going to win every job. This is the whole multi-model thesis in a nutshell. Different tasks want different tradeoffs on latency, cost, and reasoning, and the smart move was never a single super-model that does it all. It's matching the right model to each task and switching freely as better ones show up. It's the same bet we've made at Kilo from day one: model-agnostic from the ground up, 500+ models, switch mid-task, pay per token at the provider's own price. When even OpenAI is routing between models inside their own product, picking the best model for the job stops being a preference and starts becoming the architecture everyone lands on.

Beijing is discussing export controls on its top AI models — open weights included. The US already cut foreign access to frontier models once this year. Meta and Alibaba took their flagships closed. Europe runs on borrowed models, and the lending terms are changing. We wrote up what's still possible inside the window: fork the open weights while they're downloadable, fund Mistral like infrastructure, and build on model-agnostic harnesses so no policy flip forces a rebuild.