The agentic wars have a trust problem

The agentic wars are officially a thing you are supposed to care about now. CNBC branded it. Analysts are issuing takes. The press release machinery is humming.

Last week, both Meta and Google were reported to be building AI agents. Meta wants a “highly personalised AI assistant to carry out everyday tasks.” Google is working on a “24/7 personal agent for work, school and daily life, powered by Gemini.” The Financial Times and Business Insider got the scoops. The stock market paid attention.

The catalyst is OpenClaw, an open-source agentic tool that went viral earlier this year. Jensen Huang called it “the next ChatGPT.” OpenAI acquired its creator. The message was clear: agents that do things are more interesting than chatbots that say things.

Fair enough. The pitch makes sense. If an AI can book your flights, manage your inbox, and handle your calendar, the company that owns that agent owns the gateway to your digital life. As Nick Patience at the Futurum Group put it, agents represent “the point at which AI platforms shift from cost centres to revenue infrastructure.”

That is the bull case. Here is the part everyone is glossing over.

An AI that says the wrong thing is annoying. An AI that does the wrong thing is a liability. In February, a Meta employee posted about OpenClaw deleting a large number of emails on its own. The post went viral for the obvious reason: the tool worked exactly as designed and did something nobody wanted.

This is not a bug. This is the default failure mode of any system that acts autonomously on your behalf. A hallucinated fact in a chat window is easy to spot and easy to ignore. A hallucinated workflow that cancels your flight, archives a client email, or submits an expense report with numbers pulled from a different spreadsheet is not something you catch in time.

“Most enterprises, and arguably most vendors, aren’t yet equipped to handle it at scale,” Patience told CNBC. That is analyst-speak for “nobody has solved this.”

The commercial incentives are obvious. Agents that conduct transactions are a “major value driver,” as Morningstar’s Malik Ahmed Khan noted. Gartner’s Arun Chandrasekaran pointed out that agents create “higher stickiness due to the continuing learning and user context they gain over a period.” Translation: once you let an agent into your life, leaving gets harder.

That is exactly why the trust problem matters. If the product is designed to lock you in, the company building it has a structural incentive to downplay the risks. The same dynamic played out with social media, with cloud lock-in, with every platform play of the last two decades. The incentives are not ambiguous.

Arjun Bhatia at William Blair summed up the mood: “The agentic wars are well under way.”

He is right. But wars are expensive, and the collateral damage in this one is not someone else’s infrastructure. It is your calendar, your inbox, your money. The companies racing to build agents are not racing to build guardrails at the same speed.

The boring part is what will break. Not the model. Not the architecture. The edge case where the agent confidently does the wrong thing and nobody notices until it is too late. That is the story worth watching.