Harvard: AI Coding Agents Ship 30% More Code, but Software Output Stays Flat

A new Harvard study by Fiona Chen and James Stratton analyzed work data covering more than 700,000 employees at 718 companies from January 2021 to March 2026, according to German outlet t3n. Semi-autonomous AI agents lifted lines of code by 30%, commits by 20% and pull requests by 23%. Reactive AI assistants produced only a slight increase.

The catch is what happened next. In Ars Technica’s summary of the study, relayed by AI Weekly, shipped software output stayed flat and gains above 12% were statistically ruled out. Review time per pull request rose 49% and revisions nearly doubled, so human reviewers appear to be absorbing the extra output.

A separate NBER working paper found a similar pattern: big gains in commits shrink sharply by the time they turn into actual releases. For developers, the lesson is that writing code faster is not the same as shipping faster, and the bottleneck is moving to review. If you use tools like Cursor, see our Cursor AI review for how its agent mode handles this.

Meta Launches Muse Wonder With “Active Memory” and “Active Focus” for AI Agents

On October 12, Meta released Muse Wonder and Muse Wonder Full, which it calls its most capable models, built for personal and professional AI agents. Meta says both reach state-of-the-art results on agent, memory and coding benchmarks, including Harvey’s legal benchmark LAB, FocusBench and SWE-Marathon. These are Meta’s own claims.

Active Memory carries context about you across conversations. Active Focus keeps the model working toward a hard goal, trying alternatives, asking for clarification and checking in before important actions. Muse Wonder is available now through the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. Meta says Wonder will soon power its Muse assistant, starting with subscribers.

AI Weekly reports that Artificial Analysis ranked Wonder Full second on its Intelligence Index, behind Claude Opus 5.5. I couldn’t independently confirm that score, so treat the ranking as unverified for now. For the consumer side of Meta’s agent push, read how Muse compares with OpenAI’s always-on agent in our Dots by OpenAI review.

Kroger Plans to Cut 3,000 Corporate Jobs as It Leans on AI

Reuters reports that Kroger CEO Greg Foran told staff in a memo that the grocer plans to cut about 3,000 corporate roles over three years as it brings AI into its workflows. That is roughly 37% of its 8,200 corporate employees. Foran became CEO in February, and details on the cuts will be shared in the coming months.

Foran wrote that the cuts would reduce duplicative work and help Kroger lower prices and invest in in-store employees. The company cut about 1,000 corporate roles last June, and last month it outlined an AI strategy that includes models managing workflows and an AI shopping assistant. It is also under pressure to cut prices to compete with Walmart and Aldi.

Nvidia Opens Its Vera Rubin Rack Designs to the Open Compute Project

At the Open Compute Project Global Summit in San Jose on October 12, Nvidia contributed Vera Rubin rack and compute tray designs to OCP through its MGX architecture. The updated MGX 1.2 specification adds rack-scale blind-mate liquid cooling and a liquid-cooled 5,000A busbar, and supports racks up to 250kW.

Nvidia also joined AMD, Google and Meta to publish Open Data Center Specification v0.7, and Google, Meta, Microsoft and Nvidia published a grid resilience spec aimed at restoring near-full load in under two seconds after a disturbance. Nvidia released four open-source software components for its DSX AI Factory platform. Meta said it is adopting Nvidia’s Spectrum-6 Ethernet switches for factories of more than 100,000 GPUs, and Nvidia and Cisco announced a new optics form factor for 400Gb/s per lane. The point of the open standards is to give equipment makers a common target instead of waiting on each customer’s proprietary specs.

AI’s Safety Referees Are Nonprofits: METR, Apollo and Transluce Step Into the Spotlight

CNBC reported on October 11 that a small group of nonprofit evaluators, METR, Apollo Research and Transluce, is moving to the center of the AI industry as Anthropic and OpenAI lean more on third-party red teams. I couldn’t open the CNBC article directly, so these details come from AI Weekly’s summary of it.

According to that summary, the groups are acting as de facto gatekeepers because the federal government has publicly opposed creating its own regulatory evaluation body. They are still finding their footing against hyperscale spending and a fast release cadence. One caveat worth knowing: METR is a privately funded nonprofit with no statutory authority, so it can’t force a lab to delay or withdraw a model.