In early May, Anthropic revealed a new product feature for their Claude Managed Agents: dreaming. It made headlines in a big way, but most of the commentary missed the biggest takeaway of all: dreaming might just bridge the trust gap that’s kneecapped enterprise AI adoption for years.
Let’s start with what it is. Dreaming is a system that runs in the background when a Claude agent is not actively working. It reviews past experiences, identifies patterns, and discards any information or lessons that it doesn’t need any more.
It’s transformative because it allows agents to build memory over time and strip away the information that’s not relevant itself. Instead of starting from scratch with each new session, agents can now draw on every relevant experience they’ve ever had.
Think what your basic LLM can do with very little context and no memory layer. Now imagine how efficient and accurate AI outputs become when agents can lean on context and learnings they’ve accumulated over six months, or even multiple years.
Harvey, the legal AI software giant, doesn’t have to imagine. When they deployed Dreaming, they saw a sixfold increase in the completion rate for one of their legal scenarios. It’s a monumental development, and with results like that, it’s no wonder reporters wanted to shout about it. It’s also not particularly surprising that it prompted some disquiet.
The idea of AI agents that get better on their own and independently choose which pieces of information to keep sends shivers down the spines of most boards. Their nerves about AI spinning out of control and moving in directions they’d rather it didn’t is one of the biggest factors undermining enterprise AI adoption right now. In fact, recent polling from Writer shows that 35% of executives already believe they couldn’t “pull the plug” on a rogue AI agent, so it tracks that Anthropic’s new feature would concern them.
But what’s so far been missing in the conversation is that, rather than stripping away human control and sending agents off to learn without any oversight, dreaming actually puts a human in the loop.
The human user can review the memory store that dreaming has created before it goes live and starts to influence the agent’s behavior. That fact has massively flown under the radar, but it’s one of the most significant parts of Anthropic’s big reveal. The new playbook doesn’t become the live playbook until someone in your organization says yes.
It puts humans in the driving seat, which is something businesses, from leaders through to employees, have been craving for years. Because while the buzz around AI, and businesses’ desire to implement it, has remained strong since ChatGPT burst onto the scene, the lack of human oversight in many of these tools has eroded trust in the technology over time.
The truth is that employee after employee has been introduced to a new AI platform, told it’s going to transform their day-to-day, and left bitterly disappointed when it doesn’t live up to expectations.
With every new tool they’re told to use, their trust in AI tools falls away a little more. And now we’re at a point where some feel completely hopeless about using the technology in the workplace. You can see how bad this has become by looking at the attitudes of the youngest, most digitally savvy members of the workforce, who should be the most AI-friendly of us all. Gallup’s recent polling of these individuals reveals trust in AI-assisted work has dropped to just 28% in 2026, while trust in human-only work has climbed to 69%.
You can also see this erosion of trust reflected in AI adoption rates which, according to US Census Bureau data fell by a percentage point in late 2025. They dropped sharpest at larger companies (those employing over 250 people), which is not a surprise considering these bigger firms have deeper pockets and have been faster in rolling out AI tools that don’t necessarily keep a human in control.
This is why Anthropic’s dreaming is such an important milestone. Over the last year or so, the momentum has been moving in the wrong direction. AI disappointment and board’s nerves have put a dampener on adoption and the technology’s potential has not been fully realized.
But this is a firm departure point. Dreaming makes AI agents much more competent while putting human oversight front and center. Business leaders can give their employees tools that actually make their lives easier, and at the same time give them a stake in how the AI develops.
It’s not about whether a human is in the loop anymore, it’s about which human is in the loop. And businesses must now decide which individual they deem responsible for reviewing AI agents’ memory development. That’s the next big question.
For me, dreaming is the missing piece of the puzzle that will give boards and C-suites the AI confidence they’ve been waiting for. Their absence of trust has been a blocker on enterprise AI adoption for far too long, but with dreaming, the floodgates may have just opened.
