Claude Code introduces Agent View for parallel sessions and skills dispatch

• Anthropic opened Claude Code's new agent view as a research preview, with the Claude announcement repost describing it as "one list of all your sessions" and ClaudeDevs' launch thread framing it as a way to run more sessions in parallel with less overhead.
• ClaudeDevs said the new flow turns repeatable workflows into skills you can dispatch across many sessions, then step in only when a skill decides it needs human input.
• trq212's walkthrough and _catwu's setup note both describe claude agents as a terminal control plane that can sit above multiple repos and register existing CLI sessions into one pane.
• The strongest user read so far comes from Boris Cherny's post, which calls agent view the jump from one agent to many, while Aakash Gupta's thread argues the real unlock is reducing the human cost of tab-hopping rather than increasing model capability.
• The launch also lands next to a small wave of visibility fixes, with Boris Cherny's reply promising better debug logs for hangs and om_patel5's screenshot showing Claude Desktop now exposes context-window usage down to system tools, MCP tools, and the autocompact buffer.

You can watch the official demo video, skim the launch thread, and see power users already comparing it to tmux in trq212's post. The surrounding maker scene moved just as fast: a Pokemon-themed orchestration dashboard and a live diff viewer both target the same pain point, which is that multi-agent coding got useful before it got legible.

The official pitch is compact: one list of sessions, available now as a research preview. ClaudeDevs' thread adds the more important detail, agent view is meant to reduce what you have to keep in your head while several Claude Code runs are active.

That reframes Claude Code from a single interactive terminal into a supervision surface. Boris Cherny's post describes it as the step from one agent to many, and trq212's description calls it "tmux built for CC," which is probably the shortest accurate mental model.

The launch thread breaks the workflow into three pieces:

• Turn common workflows into skills, per ClaudeDevs.
• Dispatch those skills across many sessions, again from the same thread.
• Let Claude continue until the skill reaches a stop point and asks for input, according to ClaudeDevs' wording.

The notable part is not just parallel execution. It is the handoff rule. Anthropic is pushing users toward codified workflows that know when to escalate, which makes agent view look less like a prettier session list and more like a lightweight operations layer for recurring dev work.

The early usage pattern is surprisingly specific. trq212 says to launch claude agents from a high-level directory that contains all your repos, while _catwu describes hitting the left arrow from any CLI session to register that session with the control plane.

That gives the feature a concrete shape:

• One top-level terminal window runs claude agents, per _catwu.
• Existing CLI sessions can attach into that view, according to the same post.
• The pane tracks which sessions need attention and makes them easy to resume, per trq212's walkthrough.

This is a small interface shift with a big ergonomic payoff. Before this, parallel Claude Code work mostly meant terminal tabs, tmux panes, or remembering which window was still busy.

The sharpest outside read comes from Aakash Gupta's thread, which argues that Anthropic did not remove a model constraint, it removed the terminal tab as the coordination limit. His examples are managerial on purpose: bug triage in one lane, test coverage in another, refactors in several more.

That reading matches the product language better than the usual "multi-agent" hype. ClaudeDevs is not advertising smarter reasoning here. The company is advertising a way to supervise more active work without opening a transcript for every session.

The community had already started building around the missing interface. In om_patel5's Pokegents thread, a maker showed a local orchestration server with persistent agent identities, MCP messaging, notifications, and session cloning, all wrapped in a Pokemon-style dashboard.

A live diff viewer for Claude Code sessions

A day later, the same account highlighted an open source tool that shows Claude Code editing files in real time. The pitch is blunt: watch additions and removals while the agent is still working, instead of discovering the blast radius after a 40-file change lands.

Together, those projects make the launch feel slightly inevitable. Power users were already inventing their own control rooms for multi-session work. Anthropic just made the native version official.

The last interesting detail is that agent view is shipping into a product that is also getting more observable in smaller ways. Peter Yang's complaint described a Claude Code session that appeared to hang for more than three minutes with little indication of what was happening; in reply, Boris Cherny said the team was working on responsiveness improvements and debug logs.

Meanwhile, om_patel5's screenshot shows Claude Desktop exposing a detailed context-window breakdown. The UI lists messages, system tools, system prompt, skills, deferred MCP tools, deferred system tools, and an autocompact buffer, with a total of 58.1k used out of 1.0M in the example.

That matters for the same reason agent view matters. As more work gets delegated to long-running sessions and reusable skills, the product needs better ways to show state, limits, and liveness instead of asking users to trust a silent terminal.