Claude Code 更新：多会话侧边栏与任务回顾卡

Aakash Gupta @aakashgupta原推发布于 2026.10.15

在 X 上打开原推

After one interruption at work, it takes the average person about 23 minutes to get back to what they were doing. Software engineers now get interrupted by their own AI all day long, and Anthropic just shipped a Claude Code update that's basically built around that number. The 23 minutes comes from Gloria Mark at UC Irvine, who shadowed office workers and timed every task switch down to the second. Most people wandered through a couple of other tasks before they found their way back to the one that got interrupted. She also found the damage depends a lot on what pulls you away, and the really expensive interruptions are the ones that drag you onto a totally unrelated topic. Which is exactly what running a bunch of AI agents feels like. You kick off five or six Claude sessions on five or six different things, and every one of them eventually stops and taps you on the shoulder because it needs a human to make a call. So look at what Anthropic shipped this week with that study in mind. The sidebar sorts your sessions by whether they need you, so you quit checking on the ones that are doing fine. When one does need you, you get a recap card that sums up what happened before you read a single line of code, and that's about the fastest way back into a task I've seen anyone build. There's a live to-do list showing Claude's plan as it works, so you're not trying to remember six plans at once. And when Claude notices extra work along the way, it offers to spin that into its own session with one click, so a random "oh, we should also fix this" doesn't hijack what you're doing. And Claude has the same problem you do, which is kind of funny. AI models get worse as their working memory fills up with stuff that has nothing to do with the task, so Anthropic built protection in that direction too. If you've got a side question, /btw sends it to its own tab and the main session keeps going without it. When Claude splits a job into parallel pieces, each piece runs in a nested session and only hands back the finished result, so the main conversation never gets cluttered with the messy middle. How many agents one person can actually run comes down to how fast they can get back into each one. Make that faster and people run more sessions, and every extra session is usage on Anthropic's meter. Pretty good reason to sweat the details on a sidebar.

这条是引用着另一条写的：「Claude Code 桌面版多会话并行更新」