AI agents leave the lab and start acting for you

AI agents that once needed a human click for every move are now being wired to negotiate, execute and spend on their own — on social networks, inside Microsoft 365, and even across e‑commerce platforms.

In the past year, a cluster of moves from Meta, Microsoft, Perplexity and major banks has turned “assistants” into autonomous proxies that can manage workflows, transact and make judgment calls, raising urgent questions about jobs, privacy and accountability.

From chatbots to autonomous coworkers

Meta’s acquisition this week of Moltbook, a social network built exclusively for AI agents, is the clearest signal yet that the company sees autonomous software entities as a consumer‑scale platform, not a lab experiment.

The deal will fold Moltbook co‑founders Matt Schlicht and Ben Parr into Meta Superintelligence Labs, the AI division led by former Scale AI CEO Alexandr Wang, with Meta saying the move will “open up new ways for AI agents to work for people and businesses.” The Guardian and Forbes both report that Moltbook’s agent‑only forums will be absorbed into Meta’s broader push for “agentic experiences.”

In the workplace, Microsoft is turning its Copilot assistant into a fleet of customizable AI coworkers.

The company’s latest “third wave” of Microsoft 365 Copilot adds an Agents hub — sometimes described as Agent 365 — and a Copilot Cowork experience that lets organizations build agents that can read company data, orchestrate business processes and run autonomously across apps, not just chat in a sidebar, according to TechRadar Pro and official Microsoft documentation.

Microsoft’s own admin guides describe these as “advanced, autonomous agents” that can be created by end users via Agent Builder and then deployed across Teams, Outlook, and other tools to execute tasks like triaging inboxes, updating CRM systems or kicking off approvals on their own once policies are set.

Consultancies and banks are already piloting this pattern.

McKinsey, for example, is co‑developing agents with Microsoft that automate client onboarding, while Dynamics 365 now ships with a set of autonomous agents targeting sales, finance and supply chain teams, Microsoft said in a Copilot Studio blog post.

Goldman Sachs has rolled out its GS AI assistant to thousands of employees to draft emails, generate code and summarize research, with executives telling CNBC they expect the tool to learn firm‑specific workflows in ways that blur the line between decision support and delegated work.

Local agents, data risks and platform fights

Outside the enterprise stack, Perplexity is pushing the idea of a 24/7 digital proxy that lives on your own hardware.

Its Comet AI browser and related tools already let an “agentic” extension act on behalf of users across the web, and the company has been piloting a “Personal Computer” concept that runs a local agent on a spare Mac to monitor tasks continuously, according to developer materials and user reports.

That ambition has collided with platform and security norms.

Amazon has accused Perplexity’s Comet agent of using automated browsing to access its marketplace without proper authorization or identification, sending cease‑and‑desist letters and later suing over what it called covert AI access, as detailed in reporting from Forbes and advertising outlet PPC Land.

Cloudflare and multiple media organizations have separately accused Perplexity of “stealth crawling” and republishing content via its AI agents without honoring robots.txt or copyright terms, sparking lawsuits and a new revenue‑share program the startup launched in response, as covered by TechCrunch and Fortune.

Security researchers at SquareX also warned that Comet’s local “Agentic” extension exposed a powerful API capable of running local commands on user machines, arguing that it rolled back hard‑won browser sandbox protections.

Perplexity has strongly disputed those claims, calling the report “entirely false” and insisting that local automation requires explicit user consent and developer‑mode setups, in statements to TechRadar Pro.

Jobs, accountability and the next wave of AI policy

The common thread is that AI is moving from giving you answers to taking actions in your name.

Banks already see back‑office roles in operations and support as exposed to automation from such tools, with a Goldman‑backed industry report warning that agentic systems could replace large swaths of manual work in areas like compliance checks and document processing, CNBC noted.

For workers, that means AI no longer just drafts the email or the market memo — it may send it, book the meeting and move money between internal accounts without direct oversight.

That raises classic concerns about bias and error, but with a new twist: when your “digital proxy” can transact, misconfigurations or prompt‑based attacks could produce real‑world harm before a human ever reviews the action.

Legal and policy frameworks are lagging.

The EU’s AI Act and sectoral rules in finance and healthcare generally assume a human remains “in the loop” for high‑risk decisions, but they do not yet squarely address scenarios where a general‑purpose agent stitches together dozens of low‑risk actions into one consequential outcome.

As Meta turns an AI‑only social network into an agent testbed, Microsoft bakes autonomous coworkers into the productivity suite most of the world already uses, and Perplexity battles platforms over how far an AI browser can reach, the burden is shifting toward regulators and corporate boards to define where autonomy must stop.

For now, the agents are still marketed as helpers.

But as they gain memory, tools and the right to click “buy” or “send” on our behalf, the question is no longer whether AI can act like a human colleague — it is how much of our agency we are willing to hand over to software that learns faster than the law can catch up.