AI2's new coding-agent models; OpenClaw's wild ride

Welcome to Runtime! Today: AI2 introduces open-source models for building custom coding agents, the rapid rise of Clawd/Moltbot/OpenClaw, and the quote of the week.

Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime each week, and if you value independent enterprise tech journalism, click the button below and become a Runtime supporter today.

Ship it

DIY agents: As software developers quickly get accustomed to using coding agents in their process, lots of companies would like to train those agents on their own codebase but lack the resources or skills to pull that off. AI2 released a new series of open-source models this week called Open Coding Agents that use a unique technique for training those agents.

"Our method makes it easy — reproducing the performance of the previously best open-source model costs ~$400 of compute, or up to $12,000 for performance that rivals the best industry models of the same size," AI2 said in a blog post. The two SERA (Soft-verified Efficient Repository Agents) models released this week are based on the insight that the usual method for training coding agents — teaching them how to take an incorrect code sample and fix it  — can actually generate useful results even if that output is only partially correct, which means companies don't have to waste compute cycles verifying that the "corrected" samples really work.

Up, up, and away: It was a rough week for established enterprise software companies after Wall Street decided that they were all going to get killed by AI next week, or something. Airtable is still working its way up that ladder with $1.4 billion in venture-capital funding, and this week it released a new product designed to orchestrate teams of AI agents working on corporate research tasks.

SuperAgent is the no-code application builder's "first standalone product built on multi-agent coordination," Airtable said in a blog post. It was designed to take natural-language assignments from employees such as "evaluate Google as a 3-year investment opportunity" and send multiple agents off to collect data, analyze it, and deliver it in a presentation-ready format.

The Iron Claw: It was a good week for Peter Steinberger, who watched an open-source project he developed last November for running AI agents from your desktop go from little-noticed to 2 million visitors in seven days. The project was originally called Clawd, but after Anthropic "politely asked us to reconsider" Steinberger changed the name to Moltbot before finally settling on OpenClaw Thursday.

"OpenClaw is an open agent platform that runs on your machine and works from the chat apps you already use," such as WhatsApp or Slack, to execute tasks like booking calendar appointments or planning travel, Steinberger said in a blog post. Thursday's release adds support for Google Chat and Twitch as well as several security patches. 

Three is a magic number: Arcee AI is also working on open-source AI models, which have yet to reach escape velocity in the enterprise but could be the long-term strategy everyone settles on once the hype dies down and OpenAI runs out of benefactors. This week it released Trinity Large, which the company said "match[es] and exceed[es] our peers in open-base models across a wide range of benchmarks, including math, coding, scientific reasoning, and raw knowledge absorption" in a blog post.

There's been a lot of quibbling over what it means to be open source in the AI era, but Arcee AI released Trinity Large under the permissive Apache license and shared not only the weights but the training data behind the model. “Ultimately, the winners of this game, and the only way to really win over the usage, is to have the best open-weight model,” CTO Lucas Atkins told TechCrunch.

Get in formaetion: Platform Engineering Labs made its debut on Product Saturday last October with the launch of formae, an open-source infrastructure-as-code tool designed to help operations engineers manage complex distributed environments. But at launch formae only supported AWS, which it rectified this week with a new release that supports all the major hyperscalers.

That includes Microsoft Azure, Google Cloud, Oracle, and France's OVH. "This release closes a gap we kept seeing in real Infrastructure as Code setups: not the lack of integrations, but the cost and friction of extending them safely when reality doesn’t match what’s already available," the company said in a blog post.

Stat of the week

Most enterprises building apps and agents around AI models have come to the conclusion that no one model can fit every need, according to new research from Databricks. Among companies that have successfully deployed AI apps, 77% are using two or more models as the backbone for those apps.

Quote of the week

"While we believe AI labs like Anthropic and OpenAI will play a critical role in bringing AI into the enterprise, we believe the fears of broad disruption to software vendors are largely overstated." — William Blair analyst Arjun Bhatia, expressing a Runtime-endorsed view in a research note this week.

The Runtime roundup

"We found that using AI assistance led to a statistically significant decrease in mastery," according to a new report from Anthropic, which is cool.

Google Cloud's customer experience head, Hayete Gallot, has left the company after less than a year, according to Business Insider.

Thanks for reading — see you Tuesday!