• Meta tested smaller AI teams as software changes rose 220%.
• Meta saw more incidents, while specialised AI agents performed better.

Meta’s effort to reorganise parts of its workforce around smaller, AI-supported teams coincided with a sharp increase in software changes, but a smaller increase in changes that resulted in user-facing features, according to internal company data reviewed by Reuters. Internal material also showed reliability and security problems as employees increased their use of AI coding tools.

The changes were part of an internal restructuring programme known as Project OT, or Organization Transformation. Reuters reported that Meta considered scenarios in which some teams would be reduced by as much as 60%, although Meta said the figure did not apply to its overall workforce and included layoffs, redeployments, and closing open positions.

Meta also tested an operating model built around smaller teams using AI to handle work previously divided among larger groups of specialists. An early pilot involved five technology teams, each consisting of two or three engineers and one designer using AI tools.

Instead of working through established six-month product planning cycles, the groups were expected to develop prototypes in four-week sprints. Internal planning material compared traditional product teams of around 10 to 20 people with AI-supported teams of three to five.

The proposed structure typically included three or four general-purpose “builders” and one direction lead, while specialists such as designers, data scientists, researchers, and machine-learning engineers could work across several teams.

The model also reduced some management layers. Unit heads could oversee 30 to 50 people, while pod leads directed day-to-day work without necessarily becoming formal managers. By June, at least 11 engineering and research units had adopted versions of the smaller pod structure.

Meta carried out a workforce reduction in May but dropped plans for a second company-wide restructuring later in the year. Meta said around 8,000 employees were affected by the May cuts and recorded $1.18 billion in related severance expenses.

As of June 30, Meta had 75,472 employees, down 1% from a year earlier. The company said most employees affected by the May reduction would no longer be included in its reported headcount by the end of the third quarter.

Meta is continuing to spend heavily on the infrastructure supporting its AI work. The company expects capital expenditure of between $130 billion and $145 billion in 2026.

More code, more reliability pressure

Internal data showed that changes to Meta’s software platforms and infrastructure increased 220% year-on-year as employees used more AI tools. Changes that resulted in new or upgraded features reaching users rose 36% over the same period.

The figures measure different stages of software development and should not be treated as a direct productivity comparison.

Internal posts reviewed by Reuters showed that major technical and security incidents increased 40% from a year earlier, while the amount of time employees spent responding to them rose 70%.

One internal post described “reliability warning signs” related to increased AI coding as early as March. Another warned that unchecked AI agents could carry out large-scale actions that human engineers would be less likely to perform manually.

The reported incidents included service disruptions and possible data leaks. Reuters also reported that attackers exploited an AI-powered Meta customer-support bot in June to gain access to high-profile Instagram accounts.

Meta declined to comment on the internal data about AI-related disruptions. Zuckerberg later told employees that AI agent technology had not progressed as quickly as he had expected, according to the report.

Meta’s public engineering work shows that the company has also been developing controls around faster software development. In an April engineering discussion, Meta described controlled rollouts, health checks, monitoring, and regression detection as safeguards for managing changes across large-scale systems.

Meta reports gains from specialised agents

Meta has separately reported productivity gains from AI agents used in more narrowly defined engineering workflows. One example is its Ranking Engineer Agent, or REA, which is used for machine-learning experiments in Meta’s advertising infrastructure.

Meta said REA can generate hypotheses, start training jobs, investigate failures, and run repeated experiments over periods lasting days or weeks. Engineers remain responsible for strategic decisions and final approvals.

In an early production deployment, Meta said three engineers using REA generated improvement proposals covering eight models. The company compared this with a historical staffing level of about two engineers per model and described the result as a fivefold increase in engineering output by that measure.

Meta also said REA-driven iterations doubled average model accuracy over its baseline across six models. The figures were reported by Meta and have not been independently verified.

Another Meta engineering project found that agents initially struggled when working across a proprietary data-processing system containing more than 4,100 files across multiple repositories and three programming languages. The company said agents lacked information about internal dependencies, conventions, and system behaviour, and could produce code that compiled but was still incorrect.

Meta responded by using more than 50 specialised agents to analyse the codebase and create 59 context files containing information that had previously existed largely as institutional knowledge. Preliminary tests across six tasks found that the additional context reduced agent tool calls and token use by around 40%, according to the company.

The project provides a separate example of the technical work required to deploy agents inside large proprietary environments, where model training data does not necessarily contain the organisational knowledge needed to complete tasks correctly.

Independent research has also produced varied results on AI-assisted software development. Google Cloud’s 2025 DORA study, based on nearly 5,000 technology professionals, found AI adoption was associated with higher software delivery throughput but lower delivery stability, while around 30% of respondents reported little or no trust in AI-generated code.

The DORA findings distinguish between producing software more quickly and maintaining stability as those changes move through the delivery process.

A controlled 2025 study by research organisation METR found experienced open-source developers took 19% longer to complete assigned tasks when using the AI tools available at the time. METR later tested newer systems and found indications of faster development, but said selection effects and measurement problems prevented it from reliably estimating the size of the improvement.

Meta studied AI work practices in Asia

Meta also looked outside its US operations while developing its approach to smaller AI-supported teams. Reuters reported that Chief Data Officer Alex Schultz and Head of Product Naomi Gleit were among executives who visited Asia to examine how AI-focused startups organised their work.

Meta also commissioned internal research into organisational structures used by companies built around AI from an early stage. Gleit said practices she encountered while spending time at Meta’s Singapore office later influenced some teams in California and New York.

She described some of those methods as practices employees were already experimenting with rather than changes imposed entirely by senior management. The report did not identify the Asian startups studied by Meta or say that Project OT originated in Singapore.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.