Meta’s AI Agent Progress Slower Than Expected, Yet Zuckerberg Remains Optimistic

In an internal town‑hall, Mark Zuckerberg said that over the past four months the trajectory of AI‑agent capabilities has not accelerated as the company expected, even after a major restructuring.

Meta plans to spend up to $145 billion on AI infrastructure this year, part of a reorganization that included about 8,000 layoffs in May and the reassignment of thousands of employees to AI‑focused teams.

Although executives were initially very optimistic about coding tools such as Anthropic’s Claude Code, Zuckerberg admitted that the investments have “not yet been realized” and that the restructuring has not proceeded as smoothly as hoped, partly because Meta struggled to adapt quickly.

Chief Technology Officer Andrew Bosworth later clarified that an internal review found Meta did not use data from a paused mouse‑tracking and keystroke‑monitoring tool to train AI models. The tool, part of the “model‑capability program” launched in April, was suspended after employee backlash and may resume only on a voluntary‑opt‑in basis.

Engineers in the AI department report that the work environment has become extremely harsh since staff from other divisions were transferred in.

The article identifies structural engineering limits that separate prototype agents from production‑ready agents: large‑language‑model reasoning combined with deterministic tool use (API calls, code execution, database queries) works well in demos but suffers under sustained load, with context‑window degradation, inconsistent tool‑call patterns across concurrent users, and error accumulation across multi‑step task chains.

According to research from Gartner, McKinsey and Digital Applied, about 79 % of enterprises using AI agents are still in experimental or pilot phases, and only 11 % have agents operating in production. Gartner predicts that by the end of 2027 more than 40 % of AI‑agent projects will be abandoned due to rising costs, unclear ROI, and insufficient governance infrastructure.

Meta unveiled the Meta Business Agent for enterprises on June 3 in London, rolling it out globally via WhatsApp, Messenger and Instagram. The platform opened to developer partners on July 1, with billing to start on August 1, but its ability to deliver reliable autonomous workflows remains unproven.

Competitors are addressing the production gap by embedding engineering talent directly within client organizations. On June 30, AWS announced a new unit backed by $10 billion to deploy thousands of engineers on‑site, aiming to cut deployment time from months to days. Two days later, Microsoft launched Microsoft Frontier Company with $25 billion and roughly 6,000 engineers for the same purpose. Anthropic and OpenAI have launched similar initiatives. This “frontier engineering” model, first pioneered by Palantir two decades ago, is now the industry standard for bridging the gap described by Zuckerberg.

Zuckerberg’s remarks underscore that internal restructuring alone cannot accelerate the transition from model capability to production workflows; substantial on‑site engineering support is required, as demonstrated by the massive investments from AWS and Microsoft.

Meta will report its second‑quarter earnings at the end of the month, giving investors a chance to question the disparity between capital expenditure and actual progress, and to assess whether the company can meet its guidance for the remaining fiscal year.