New ‘Agent View’ dashboard streams Claude Managed Agents’ session activity

The thing Anthropic shipped is not a smarter model. It’s a control surface. Agent View, released on May 11 inside Claude Code, gives developers one screen for all the Claude sessions they have running at once. That sounds small, but it fixes a real bottleneck. The problem with agents is no longer just getting one to do a task. It’s managing five, ten, or more without losing track of what each one is doing. Anthropic’s answer is a dashboard in the CLI that turns a pile of background sessions into something you can actually supervise. ### What did Anthropic actually launch? Agent View is a research-preview feature in Claude Code. You open it with `claude agents`, and it shows a table of your sessions — what’s running, what’s waiting on you, and what’s finished. You can start new sessions from the same screen, peek into the latest turn, reply inline, or attach to the full conversation when you want to take over directly. ### Why is that useful now? (claude.com) Because developers have moved past the “one agent, one terminal” stage. Anthropic’s own pitch is basically: stop juggling terminal tabs, tmux panes, and memory. If you’re running a bug fix, a PR review, a log investigation, and a long-running maintenance task in parallel, the hard part becomes orchestration. Agent View is meant to be the missing dispatcher. ### What makes this more than a prettier list? (claude.com) The background execution model. Sessions keep running even when no terminal is attached, and you can send an existing interactive session into the background with `/bg` or start one directly with `claude --bg "<prompt>"`. That matters because it turns Claude from a foreground chat tool into something closer to a managed worker pool. The dashboard is just the human-facing layer on top. ### How does this connect to Managed Agents? Pretty directly. In April, Anthropic described Managed Agents as a hosted service built around three abstractions: the session, the harness, and the sandbox. A session is the append-only log of everything that happened. The harness runs the loop and routes tool calls. The sandbox is where code runs and files get edited. Agent View feels like the UI expression of that architecture — especially the session part — even though this launch is framed inside Claude Code. (code.claude.com) That last sentence is an inference, but it lines up with how Anthropic has described the system. ### Why does “session” matter so much? Because long-running agents don’t fit neatly inside one context window. Anthropic has been building toward systems that can carry work across resets, tools, and multiple turns without collapsing. In its autonomy research, the company said long Claude Code sessions among the heaviest users nearly doubled in three months, from under 25 minutes to over 45 minutes. Once agents run that long — and once you run several at once — visibility stops being a nice extra and becomes core infrastructure. (anthropic.com) ### Is this also about evaluation? Indirectly, yes. Anthropic has spent the last year pushing the idea that agents need better grading, not just better prompting. Its evals work argues that agent systems need explicit tasks, graders, and repeated trials because errors compound over many turns. Agent View doesn’t grade anything by itself in the launch materials, but it does make the raw behavior legible — what the agent did, where it paused, and when a human had to step in. (anthropic.com) That’s the stuff teams need before they can improve reliability. ### Why is Anthropic pushing this now? Because the company is clearly leaning into agents as a product category, not a side feature. In the last week alone, Anthropic also released ten finance-focused agent templates that can run as cookbooks for Claude Managed Agents. More domains, more templates, and more autonomous runs all create the same need: supervision that scales with concurrency. ### Bottom line (anthropic.com) Agent View is basically Anthropic admitting that the next problem in AI agents is management. Model quality still matters. But once agents can work in parallel for long stretches, the winning product also needs a cockpit. (claude.com) (anthropic.com)