--- title: "Claude Code v2.1.139: Agent View Turns Your Terminal Into a Fleet Dashboard" date: 2026-05-20 tags: ["claude-code","anthropic","agent-view","agentic-coding","background-sessions","goal-command"] categories: ["AI Tools","Agentic Workflows"] summary: "Claude Code v2.1.139 ships two features that change how multi-agent work actually looks: Agent View — a unified dashboard showing every running, blocked, and completed session — and the /goal command, which keeps Claude working autonomously across turns until a defined completion condition holds." --- Claude Code has supported background sessions and parallel worktrees for months. The workflow was real but the visibility was not: you launched agents with `claude --bg`, context-switched between terminals to check progress, and manually tracked which sessions were waiting on you versus still running. v2.1.139 closes that gap with two features that make multi-agent development something you can actually see and direct. ## Agent View: One Screen for Every Session `claude agents` opens a live list of every Claude Code session in the current environment, organized by state: - **Running** — the agent is actively working on its current turn - **Blocked** — the agent needs a human decision before it can continue - **Done** — the agent has completed its work and is ready for review The view updates in real time. If you have five background sessions across three repos and one is waiting on a tool approval, it appears in the Blocked row immediately. You do not have to cycle through terminal tabs or check process output to discover that. The practical effect is a shift in how you allocate attention during multi-agent work. Today's pattern: launch a task, forget it, discover it was stuck ten minutes ago when you check back. The Agent View pattern: glance at the dashboard every few minutes, handle anything in the Blocked column, continue with whatever you're doing. Agent View also shows sessions started via `claude --bg` alongside interactive ones, with background sessions marked `bg`. The `/resume` command works directly from the view, which means handling a blocked session no longer requires remembering which background session handles which task. The feature ships as a research preview, which is the same stage Computer Use and Code Review shipped before going GA. Expect it to evolve based on usage patterns before it stabilizes. ## /goal: The Command That Doesn't Stop The /goal command takes outcome-based execution seriously. Instead of running a task and returning control when it completes, /goal defines a completion condition and Claude keeps working across as many turns as necessary until that condition holds: ``` /goal All tests in the auth module pass and coverage is above 85% ``` After each turn, a separate Haiku model evaluates whether the condition holds. If the condition is not met, Claude starts another turn. If it is met, execution stops and control returns to you. The dual-model design is worth paying attention to. Using the same model to decide both what to do and when to stop creates a failure mode where the agent convinces itself the condition has been met before it actually has — what practitioners call mission drift. By running the evaluation in a separate Haiku context that receives only the current state and the original condition, the architecture keeps the judgment about success structurally separate from the work itself. While /goal runs, a live overlay panel shows elapsed time, turn count, and current token usage. You can monitor cost accumulation in real time rather than discovering a large bill after the fact. The command works in interactive mode, with the `-p` flag for scripted invocations, and in Remote Control. This means you can wire /goal into a CI pipeline, a Routine, or any other orchestration layer that drives Claude Code programmatically. ## What Else Shipped in 2.1.139 The two headline features are accompanied by a set of smaller but useful changes: **Plugin improvements**: `claude plugin details` now shows a plugin's full component inventory and a projected per-session token cost estimate before you install or run it. The marketplace also enforces plugin dependencies — if a plugin requires another plugin or MCP server to function, the dependency is flagged at install time rather than discovered at runtime. Both changes address the common experience of installing a plugin that silently fails to work. **Session and background controls**: Background sessions (`--bg`) now appear in interactive mode session lists marked as `bg`. The `/resume` command accepts session IDs from background sessions, making it consistent with the existing interactive session resumption flow. **Transcript navigation**: The transcript view now supports keyboard shortcuts for jumping between user prompts. For long sessions involving many turns, this is meaningfully faster than scrolling. **Observability for multi-agent work**: API requests from subagents now carry `x-claude-code-agent-id` and `x-claude-code-parent-agent-id` headers. The `claude_code.llm_request` OpenTelemetry span includes matching `agent_id` and `parent_agent_id` attributes. Teams running multiple Claude Code agents in orchestrated workflows can now trace which agent made which API call without correlating by timestamp. **/scroll-speed**: A minor quality-of-life command that adjusts mouse wheel scroll speed in the terminal with a live preview. Trivial, but the kind of thing that grates when it's wrong for your setup. **Notable fixes**: A deadlock that blocked `claude auth login`, `logout`, and `status` when expired credentials coincided with the `forceRemoteSettingsRefresh` enterprise policy is resolved. The `autoAllowBashIfSandboxed` flag now correctly approves commands that include shell expansions. Unbounded memory growth from HTTP/SSE MCP servers streaming non-protocol data is patched. ## From Session to Fleet The direction of 2.1.139 is readable: Claude Code is evolving from a per-session tool into a multi-agent coordination platform you operate from a single interface. Agent View is the control plane. /goal is the autonomy primitive that makes agents worth controlling. The observability additions (OTel agent headers, projected plugin costs) are the instrumentation layer that lets you understand what the fleet is doing and what it costs. These three things are coherent; they belong together. The practical ceiling for today's solo developer is probably eight to twelve parallel sessions before coordination overhead exceeds productivity gain. But the relevant frame is not the solo developer — it is the team running Claude Code Routines overnight, the enterprise running Code Review on every PR, the startup where one developer is directing a dozen specialized agents across a monorepo. Agent View makes that model of working possible to manage without losing track. Upgrade via `npm install -g @anthropic-ai/claude-code` or wait for the auto-update. The Agent View research preview requires no configuration beyond launching `claude agents`. --- **Sources:** - [Release v2.1.139 · anthropics/claude-code](https://github.com/anthropics/claude-code/releases/tag/v2.1.139) — GitHub - [Claude Code CLI 2.1.139 changelog](https://x.com/ClaudeCodeLog/status/2053913638197416198) — ClaudeCodeLog on X - [Claude Code v2.1.139: Agent View, Goal Setting, and Enhanced Workflow Control](https://claude-world.com/articles/claude-code-21139-release/) — ClaudeWorld - [Claude Code Agent View and Goal Command for AI Engineers](https://zenvanriel.com/ai-engineer-blog/claude-code-agent-view-goal-command-guide/) — Zen van Riel - [Claude Code 2.1.139 adds /goal command](https://explainx.ai/blog/claude-code-goal-command-long-running-agents-2026) — explainx.ai - [Changelog - Claude Code Docs](https://code.claude.com/docs/en/changelog) — Anthropic