Your AI Agent Wasn’t at Lunch
Yesterday, Anthropic introduced Claude Duet, a new personal AI agent designed to work across connected applications, keep track of what needs attention, and complete assignments even when you’re away from your computer. It joins Meta’s Muse, OpenAI’s Dot, and Instinct as part of a growing wave of personal AI agents designed to do much more than answer our questions. These systems are moving toward a future where they can follow our work and increasingly handle tasks on our behalf.
I’ve been thinking a lot about what that might look like in law offices and chambers, particularly as we connect these agents to the Microsoft Teams and Slack channels where so much of our daily work takes place. Imagine an agent that follows discussions, remembers decisions, and keeps track of assignments. A lawyer returning from court could ask what happened while she was gone, or a judge could check on a project without interrupting a staff member. Eventually, the agent might even handle some of the follow-up work itself. The possibilities are pretty impressive, but the more I think about these agents becoming part of our everyday routines, the more I wonder whether we’re overlooking how much of our work never gets written down.
Think about a typical day in an office. Someone raises an issue in Teams or Slack, a few people weigh in, and the group spends part of the morning discussing how to handle it. Later, two of those people grab lunch together and, somewhere between ordering their food and paying the check, realize the original approach isn’t going to work. They come up with a better idea, mention the change to a few colleagues when they return, and by the end of the afternoon everyone has moved on.
The next morning, one of the people who was at lunch asks the AI agent to move forward with the project. The agent reviews the messaging discussion, identifies what appears to be the agreed-upon approach, and gets to work preparing a document, sending an email, or coordinating the next steps. Unfortunately, it’s working on the wrong plan. The person who gave the agent the assignment was there for that conversation. He knew the plan had changed, but it never occurred to him that the agent didn’t.
Of course, people have always assumed colleagues knew things they didn’t, and we’ve all experienced the confusion that follows. But there’s something different about an AI agent that can read thousands of messages, remember every detail, and act on that information almost instantly. The more capable these systems become, the easier it may be to assume they have a complete understanding of what’s happening simply because they have access to our communications. An agent might have a more complete record of our written communications than anyone in the office, yet still have a less complete understanding of what’s actually happening.
As these agents move beyond answering questions and start taking action, that gap in understanding becomes more consequential. An agent could follow every instruction, accurately interpret every message, and still carry out an assignment based on a plan that everyone else has abandoned. The possibilities for these tools remain enormous, but we may need to develop some new habits as they become part of our daily work.
After all, your AI agent wasn’t at lunch, and you need to tell it what happened.
