Welcome to Runtime! Today on Product Saturday: Glean launches a new desktop app for searching across corporate documents and data, Harness wants to be the home for your AI-generated code, and Thomson Reuters makes its own model.

Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime for free every week, or level up here.

Ship it

See the light: Agent harnesses haven't caught on with the general business user to the same extent that coding agent harnesses have altered the trajectory of software development, but that's starting to change. Still, a lot of companies that are concerned about data security and runaway token costs need assurances that hooking employees up with autonomous business tools won't result in sensitive data flying out the window in the most expensive way possible.

Glean has raised $768 million to build out its enterprise search platform, which uses agents and coworking tools to allow IT departments to trust their users (to the extent any IT department ever does, anyway) with autonomous workflows. This week it announced a new version of the desktop experience of that platform called Glean Tau, which it compared to Anthropic's Claude Cowork.

Glean Tau "can plan, execute, review, recover, and act on a user’s behalf across tasks like organizing files, analyzing documents, creating spreadsheets, and working across connected apps," the company said in a press release. It hooks into Glean's broader enterprise platform, which makes it easier for companies to put restrictions on data access without preventing employees from finding the right data for the right task.

But Glean's main message behind the introduction of Glean Tau was to appeal to the tokenomics crowd, claiming that Glean is about 80% cheaper to use than Claude Cowork based on the results of its own benchmark, which is a little funny. Your mileage may vary, but if AI-powered business assistants are going to make a real dent in the enterprise, cost — as always — will be a big factor.

Delivery, continued

The year of the GitHub-killer: GitHub's ongoing inability to provide a stable home for enterprise code (you can read about this week's incidents here) will force some companies to make hard decisions as agents redefine almost every part of software development, and the sharks are starting to circle. Last week we highlighted the launch of Cursor's Origins service, and this week Harness rolled out its own take on a modern repository.

"Harness Code Repository is what source control looks like when agents are part of the team," Harness said in a press release. It was designed to "handle thousands of pull requests and commits opened at once," the company said, and comes with strict access-control features that users can apply to AI agents to keep them in their lane.

Moar tokens: This week's Hot Chips conference featured several interesting launches, including SiFive's first RISC-V server, but Nvidia continues to rule over the AI chip landscape. The fruits of its kinda-sorta acquisition of Groq last year were released this week, with the Groq 3 LPX inference chip reaching full production.

Designed as a co-processor with the Vera Rubin rack-scale server, the new chip's mission is to increase "the rate at which tokens are generated for an individual user, [which determines] how quickly an agent can complete each step of its work," it said in a press release. The Groq licensing deal was an interesting admission from Nvidia that while its flagship GPUs remain the state-of-the-art for training, AI inference is going to require some different approaches.

DIY AI: The rise of open-weight models has been one of the more interesting developments in enterprise tech this year, helping companies control costs, setting up a market for AI gateways that route traffic between models, and allowing some businesses to take their AI strategy in-house. Thomson Reuters published a paper this week detailing how it trained a new model called Thomson (of course) using open-weight models such as Aliababa's Qwen.

The new model, which is also open weight, "performs competitively with recent frontier models on a wide range of domains and capabilities, ranging from agentic tasks to safety, legal, tax & multilingualism, to comprehensive large-scale Deep Research," the company said in the paper. Thomson Reuters also said that it spent just $450,000 on the final training run for the large version of the model, while estimating the total cost required to develop the model at around $40 million.

Quote of the week

"When it breaks, what's going to happen is your leadership is going to try to hold you accountable for the system that you built. And if you don't understand what you built, you're going to start panicking." — Microsoft principal engineer Raki Rahman, who told The Stack that even in the AI coding agent era, software developers need to own their code.

We're also reading:

One of East Coast’s largest data centers accused of "violating federal law": Floodlight sent a drone over a prominent data-center facility in New Jersey run by Nebius, and just like Elon Musk's disregard for the health of the citizens who live around SpaceXAI's massive data center in Memphis, found the facility is operating "dozens of unpermitted gas-powered generators."

Thanks for reading — see you Tuesday!

The link has been copied!