We're teaming up with Hugging Face to make open agent traces the fuel for the next generation of open coding models. You can now upload @Droid agent traces directly to @huggingface datasets.

GLM 5.2 Fast is now available in Droid, hosted by @FireworksAI_HQ. GLM 5.2 is one of the most popular models in Droid. Now you can build with the same quality at much higher speed.

This is exactly why we believe in customization. Quick context: Factory's original secret scanner was deterministic, so it either flagged things that weren't actually secrets (false positives) or missed secrets that didn't match its patterns (false negatives). Their solution was elegant: two small post-trained models—one to catch missed secrets and another to filter out false alarms. The result: models that outperform GPT-5.5 and Opus 4.8 on this specific task while running at a fraction of the cost and latency. This is our thesis in action: take a strong open model, post-train it for a specific production problem, and you can build something that's faster, cheaper, and better than frontier models at that job. Congrats to @FactoryAI. Proud they built Droid Shield 2.0 on @FireworksAI_HQ.

This was fun! We used @FireworksAI_HQ for training two LoRA adapters on top of a small MoE model to catch more secrets and address friction. Weights are available on @huggingface for the community to freely access and continue to build on top of.

4 of the last 5 PRs I merged in the last 3 days were one-shots from a single Slack message. Slack has become the default interface for my job. We've set up enough guardrails (hundreds, from CI checks, lint rules, security scans, e2e/unit tests, automated QA) that shipping a bug is virtually impossible here. On the off chance the result is drifting off course, I can take over the live session from the app to steer it back into place. Just like FSD with a Tesla, but the default is auto-pilot. Most of the time I skip manually testing. I read the droid summary, skim the code, run the preview deployment, click around a bit, or just watch the GIF the QA bot records of itself going through the app. Last week I tagged Factory for a fix, and the PR was created, passed CI, approved, and merged in under 12 minutes. 1 min to spin up a sandbox, 3 min to code, 6 min for CI to pass and finally 2 min for review. The software factory is real, and it is here.

"We don't need to live like this." Matan Grinberg (@matanSF) is the founder and CEO of @FactoryAI, a startup building AI agents, called droids, that handle software engineering the way no single AI lab wants them to: independent of any one model. His conviction is that AI's real constraint isn't capability. It's that the industry keeps building tools locked to one company's model, leaving that company free to dictate the terms of how you can use it. (0:00) Intro (3:50) Noether’s theorem explained (10:53) Why there will always be more problems to solve (20:10) Why Factory abstracts away model choice (35:33) How Matan got into string theory (41:53) Startup founders vs. theoretical physicists (52:53) The origins of Factory (1:08:11) Matan’s predictions for the future of AI and Factory Thank you to the partners who make this possible @dottechdomains: An identity for builders at their core: @brexHQ: The intelligent finance platform: @withpersona: Trusted identity verification for any use case:

we’ll be at AI Engineer After Dark on July 1st with @vercel , @merge_api, @FactoryAI , and a room full of people building the AI engineering stack. our Research Lead, RL Training @olive_jy_song will be giving a lightning talk on post-training MiniMax M3 as part of a lineup featuring @browserbase , @neondatabase , @FireworksAI_HQ , @ExaAILabs , @sentry , @SurrealDB, and more. see you after dark at SFMOMA.