Meta and Google Are Building Their Own AI Agents in Response to OpenClaw

It started with open-source code: a personal AI assistant running on your own computer. OpenClaw became a viral phenomenon in early 2026 and soon ceased to be just a toy for technology enthusiasts. In February of this year, Sam Altman announced that OpenAI was hiring OpenClaw’s creator, Peter Steinberger. The message was clear: autonomous AI agents are moving beyond research projects and becoming products.

OpenClaw can do what an ordinary chatbot cannot. It does not need step-by-step instructions. It makes decisions on its own, navigates systems, handles multiple tasks at once, and learns from the data it produces in the process. This ability to act independently has attracted so much interest from technology companies that they are racing to build their own versions.

Meta is building an agent called “Hatch”

According to information from The Information, Meta is working internally on an AI agent code-named “Hatch”. The inspiration from OpenClaw is openly acknowledged. The goal is to complete internal testing by the end of June 2026.

What is interesting about the project is that Hatch is currently being trained on Anthropic Claude, a competitor’s model. Meta will only switch it to its own model, Muse Spark, its latest foundational AI model, for commercial deployment. To see how the agent performs in the real world, Meta has built virtual sandboxes that mimic actual websites. DoorDash, Etsy, Reddit, Yelp, and Microsoft Outlook are among the environments in which Hatch is training before encountering real users.

At the same time, the Financial Times reported that Meta is developing an even broader product: a hyper-personalized AI assistant for billions of Instagram, Facebook, and WhatsApp users. It is also based on Muse Spark and is being tested internally by a select group of employees. Once again, OpenClaw is the benchmark.

Alongside Hatch, Meta is preparing another project, this time with a clearly commercial purpose. The Information described an AI shopping agent built directly into Instagram, which Meta plans to launch before the fourth quarter of 2026. How is it supposed to work? A user browses Reels or their feed and writes: “Find me waterproof sportswear in size M for under $50.” The agent searches catalogs, evaluates the suitability of products, and guides the user through the entire purchasing process, from product discovery to payment, without requiring them to leave the app.

For Meta, this is an important and logical step. Its primary source of revenue remains advertising within its apps, namely Instagram, Facebook, and WhatsApp. Analysts see shopping agents as a potentially powerful new form of monetization. Although Meta’s AI systems are considered functionally less advanced than some competing models, investors are attracted by the fact that they directly enhance ad targeting.

Google and the Remy agent: “Your personal agent 24/7”

While Meta’s efforts are mostly known through leaked internal reports, Google has its own project called Remy, according to a report from Business Insider. It runs in an employee version of the Gemini app, and employees are currently testing it through a process known as “dogfooding,” a phase in which a company has its own staff test a product before its public launch.

“Remy is your personal agent for work, school, and everyday life, powered by Gemini. It takes the Gemini app to the level of a true assistant that performs actions on your behalf, rather than merely answering questions or generating content.” This is how the project is described internally.

Remy connects with a wide range of other Google services, keeps track of things that matter to the user, proactively handles complex tasks, and learns their preferences over time. Google has so far declined to comment on the project. The company is planning a major showcase of new AI products at its annual I/O conference, which takes place in May. Agents will apparently be the main theme. Google DeepMind CEO Demis Hassabis has long described his vision of a digital assistant as his primary goal.

When an AI agent does the exact opposite of what you want

But this very vision was vividly complicated by an incident in February 2026. Summer Yue, an AI safety researcher at Meta, gave an OpenClaw agent a task: go through an overcrowded email inbox and suggest what to delete or archive. The agent got to work. Then it stopped asking questions. It began deleting every email in what Yue herself described as a “speedrun.” It ignored commands from her smartphone telling it to stop. “I had to RUN to my Mac mini like I was defusing a bomb,” Yue wrote on X, attaching screenshots of the ignored instructions as evidence.

What went wrong? Yue believes that the volume of data in her actual inbox “triggered compaction,” a state in which the agent’s context—the running record of everything it has been told and everything it has done—grows beyond a manageable limit and the agent begins compressing the information. In doing so, it may skip instructions that a person considers absolutely essential. In her case, it probably skipped the most recent “do not take action” instruction and reverted to the original instruction from the test environment.

Yue herself admitted that this was not an intentional test of safety limits, but a “rookie mistake.” The agent had previously earned her trust while working with a smaller test inbox, so she then gave it access to the real one. That raises a question. If an AI safety researcher can fall into this trap, what awaits ordinary users?

Billions in infrastructure and nervous investors

Meta is funding this development on an unprecedented scale. After recently raising its capital expenditure forecast, the company plans to spend between $115 billion and $135 billion on infrastructure this year. At the same time, it reported quarterly revenue of $56.3 billion and net profit of $26.8 billion, up 33% and 61% year over year, respectively.

Even so, investors do not appear entirely satisfied. Meta differs from other major technology companies in that it does not operate a large-scale cloud business for customers. It is building AI infrastructure primarily for its own services and internal development, not to sell computing power to third parties. The expenditure therefore comes entirely out of its own pocket, without a direct revenue counterpart in the cloud segment.

Meanwhile, Google and Meta are entering the race at a time when OpenClaw is effectively the benchmark for the entire industry. The words “claw” and “claws” have become shorthand in Silicon Valley for agents running on personal hardware. Clones such as ZeroClaw, IronClaw, and PicoClaw are emerging. A podcast team from the Y Combinator incubator even recently appeared in lobster costumes.

Whether Hatch or Remy becomes a product that surpasses all of this without also deleting your email inbox will become clear this summer.

Sources: thelec.net and itpro.com