According to people familiar with the matter, Meta Platform Company is building an AI agent for ordinary consumers and benchmarking OpenClaw, and is also developing a new smart shopping tool. At present, Meta CEO Mark Zuckerberg is vigorously promoting the implementation of new AI products, hoping to bring actual returns to the company's huge investment in artificial intelligence.
Meta CEO Mark Zuckerberg
One of the people familiar with the matter said that Meta is training this internal project based on the OpenClaw concept, with the internal code name Hatch, and plans to start internal testing before the end of June.
The name of this agent may be changed after it is officially launched in the future. Its goal is to complete various daily tasks for users. People familiar with the matter said that during the model training process, Meta built a sandbox network environment to simulate real websites such as DoorDash, Etsy, Reddit, Yelp, and Outlook for simulation testing of Hatch.
Another person familiar with the matter revealed that Meta plans to embed an independent smart shopping tool into Instagram and aims to officially launch it before the fourth quarter of this year.
Zuckerberg has recently paid close attention to the development potential of AI agents and regards this technology as the core of his vision of “personal superintelligence”.
Meta said during last week’s earnings call that it will increase AI infrastructure capital expenditures to a maximum of $145 billion this year. Zuckerberg said at the meeting that the company’s goal is to “create AI agents that can understand user goals and autonomously assist users to achieve their goals around the clock.”
He also admitted that there are still many technical challenges in adapting tools like OpenClaw to Meta’s billions of users: it requires huge computing power infrastructure support, and it also needs to be simple and easy to use.
OpenClaw has become popular among technology enthusiasts this year. Users can use it to build independent AI agents; however, for the vast majority of non-professional ordinary users, its operating threshold is too high and too complicated.
Peter Steinberg, the developer of the tool, revealed in a podcast that Meta had intended to acquire OpenClaw earlier this year, but the AI agent was eventually acquired by OpenAI in February.
According to a third person familiar with the matter, Hatch’s current R&D training is based on Anthropic’s Claude Opus 4.6 and Claude Sonnet 4.6 large models; after it is officially launched, it will switch to Muse Spark, Meta’s latest self-developed large model.
Informed sources said that Meta is stepping up efforts to optimize Hatch’s active decision-making capabilities so that it can independently judge and perform tasks without waiting for user instructions; at the same time, it will expand the amount of information the model can process at a time, strengthen long-term memory, and retain key details in multiple rounds of conversations. In addition, Meta is still polishing the user interaction logic of the agent, as well as the tool calling and function selection mechanism.
The Financial Times has previously reported that Meta is developing a highly personalized, OpenClaw-like AI assistant; the Hatch project and this new shopping tool have been exposed for the first time.
Currently, major technology giants are competing to deploy the consumer-oriented AI intelligent industry. It is generally believed that such products will become the next generation of mainstream interaction portals for people to shop, work, and socialize online in the future.
Smart shopping agents have become a popular track for major technology companies.
In January this year, Google launched the Gemini enterprise version for customer experience. It has a built-in AI shopping agent that can recommend products, help users make additional purchases, and complete orders after confirmation by the user. Amazon has also launched its AI shopping assistant Rufus, which supports price tracking, product research and one-click purchasing.
A fourth person familiar with the matter said that Meta will embed the shopping intelligence into Instagram in order to strengthen competition with TikTok’s small store e-commerce business.
In the future, users can click on products in Instagram short videos or information streams to view details, jump to external web pages, and complete purchases directly within the platform.
This new tool will inherit Meta’s new AI shopping experience released in March this year: relying on AI to generate more detailed product information, and launching a new checkout function, users can click on the advertisement to place an order directly.
Currently, shopping agents still face multiple obstacles to implementation: some platforms refuse to open permissions and restrict access to other AI agents.
Amazon CEO Andy Jassy said last week that the overall experience of third-party AI agents on the Amazon platform is still not ideal, and problems such as pricing errors and inaccurate product information often occur.
At the same time, some AI agents have made mistakes such as providing incorrect professional advice. Meta and other technology companies still need to convince consumers to trust the reliability of such tools.
Meta has also launched the self-developed AI intelligent agent MyClaw internally for employees to retrieve work documents, summarize forum posts, and obtain technical consultation.
Previous media reports revealed that an employee had acted according to incorrect suggestions given by MyClaw, which resulted in sensitive company and user data being obtained by unauthorized personnel, triggering a major Meta security alert.
Meta acquired Manus, an AI agent project developed by a Chinese startup, as early as December last year; but China’s National Development and Reform Commission said last month that it required Meta to divest the acquired assets.