hunter-gplpcn
In an internal meeting in the first week of July, Meta executives delivered a message to employees that the company’s public statements had only hinted at: development of AI agents has not met the goals the company set at the start of the year. The executives said they expect “substantive progress” within a few months, according to people who attended.
The acknowledgment tracks a comment from Mark Zuckerberg, who said earlier this year that AI agents had not accelerated at the pace the company expected. The internal remarks were blunter. Executives said the gap between the company’s ambitions and what its agents can actually do in the real world — booking, shopping, completing multi-step tasks without human help — remains wider than planned, the people said.
The targets the company missed were set in January, when Meta mapped a year of AI products and set internal targets for agent reliability. People familiar with the plans said the models improved steadily through the spring, but the systems wrapped around them — memory, tools, payments, safety guardrails — lagged. The result is a familiar pattern: the intelligence got better faster than the scaffolding needed to let it act.
The admission comes at a delicate moment for Meta. The company has made AI the center of its story, telling investors that spending on models and infrastructure will pay off in a new wave of products. At the same time, it is running several bets at once: the Llama family of open models, an AI game called Pocket, and a push to transform its cloud business. Each line is progressing, executives said, but none has yet produced the kind of breakout that would quiet the question of whether the spending is working.
The substance of the meeting, people familiar with the discussion said, was about what “substantive progress” would look like. Executives cited specific capabilities — agents that can complete real transactions, hold longer conversations without losing context, and coordinate across the company’s apps — rather than the demo videos that dominated earlier internal presentations. The shift in language matters: it suggests the company is moving from showing what agents might do to measuring what they actually do.
Pocket, the AI game, is a test of a different kind. Games have long been a proving ground for AI systems, because they require real-time decision-making and reward continuous improvement. Meta’s bet is that a game people actually play will train its models in ways benchmarks cannot. Executives at the meeting pointed to early engagement data as a bright spot, while cautioning that the project is years from a verdict.
The gap between publicity and reality is a familiar feature of the AI cycle. Every major lab has announced agents that work flawlessly in demonstrations and stumble in production. Meta’s particular problem is that it promised a consumer-scale version of the promise: agents that live inside WhatsApp, Instagram and Facebook, where hundreds of millions of people could encounter them daily. A stumble in that setting is visible in a way that a backend failure is not.
The internal message also carries a financial subtext. Meta’s capital spending has climbed sharply, and investors have tolerated it on the assumption that AI revenue is coming. If agents are behind schedule, the timeline for that revenue slips with them. Executives did not change the company’s financial guidance at the meeting, the people said, but they acknowledged that the pace of adoption — not the pace of research — now determines when the spending turns into returns.
The competitive clock is also running. OpenAI and Google have both shipped agent products aimed at consumers, and Anthropic has made enterprise agents its focus. Meta’s advantage is distribution — its apps reach billions — but distribution only matters if the underlying agent works. Analysts said the next two quarters will be decisive: either the “substantive progress” arrives and Meta converts its user base into agent traffic, or the company faces a harder conversation with investors about the returns on its infrastructure spending.
The January targets were not modest. Meta had planned for its agents to handle a meaningful share of consumer transactions by mid-year, and to support a public beta of agent-to-agent commerce across its messaging apps. Neither has arrived on the promised schedule, the people said, though both remain active projects. The gap between the plan and the reality has consequences beyond morale: partners who built integrations around Meta’s stated timelines have had to wait, and some enterprise customers have chosen rival platforms in the meantime. Executives at the meeting said the delays reflected engineering reality, not lost interest, and pointed to the agent infrastructure now in testing as evidence that the pieces are coming together.
For employees, the meeting was a mix of candor and reassurance. The company is not cutting the program; it is pressing it. The message, several people who attended summarized, was that the technology is closer than the public believes, and that the months ahead will show it. Whether that turns out to be true is now the central question for a company whose stock price has come to move with its AI narrative.
