Claude Code’s Agent View Turns One Chat Window Into an Operations Console
Claude Code’s new Agent View looks, at first glance, like a terminal table. That is underselling it. Tables are what developer tools become when they stop being demos and start being queues.
Anthropic shipped Claude Code v2.1.139 on May 11 with a research-preview claude agents interface for supervising background sessions, plus a new /goal command, richer plugin inspection, safer hook execution, MCP reconnect improvements, and new OpenTelemetry identifiers for subagent requests. The visible feature is simple: one place to see which Claude Code sessions are working, blocked, idle, failed, stopped, loop-sleeping, or ready for review. The strategic change is larger: Claude Code is moving from “one assistant in one transcript” toward a local operations console for parallel software work.
That shift matters because the old chat-window metaphor was already buckling. Engineers do not work on one perfectly linear task. They investigate a flaky test, wait for CI, review a migration, patch docs, chase a security warning, and answer Slack in between. A single scrolling transcript is fine for a five-minute refactor. It is a bad control surface for a day full of delegated jobs.
The unit of work is becoming the job, not the conversation
The Agent View docs describe a deliberately queue-shaped interface. It opens with claude agents, requires Claude Code v2.1.139 or later, and can be disabled centrally through the disableAgentView managed setting. Sessions are grouped by operational state: working, needs input, idle, completed, failed, stopped, and loop-sleeping. Users can peek with Space, reply from the peek panel, attach with Enter or right arrow, pin sessions, reorder them, group by state or directory, filter by agent/state/PR, and dispatch new background sessions.
That is not chatbot UX. That is queue management. It is closer to watching CI jobs, Kubernetes pods, or background workers than talking to a pair programmer. The product is admitting that the human’s role is increasingly supervisory: assign the work, inspect blocked states, attach when judgment is needed, and review the artifact.
The persistence model reinforces the point. The docs say sessions persist on disk and do not need a terminal attached. A supervisor process keeps them running, and if the machine sleeps or shuts down, users can restart work with claude respawn --all. That is a meaningful boundary crossing. Once a coding assistant keeps working without a live terminal, it becomes a local runtime with recovery semantics, not just an interactive CLI.
Anthropic also added /goal, a completion condition that keeps Claude working across turns until the condition is met. It works in interactive mode, -p, and Remote Control, with an overlay showing elapsed time, turns, and tokens. This is the other half of the same design: define a job, let it run, watch progress, and intervene when needed. The interface is catching up to the workflow users were already trying to force through transcripts.
Parallel agents need boring control-plane plumbing
The most important parts of v2.1.139 are not all visible in the table. Subagent API requests now carry x-claude-code-agent-id and x-claude-code-parent-agent-id headers, and claude_code.llm_request OpenTelemetry spans include agent_id and parent_agent_id attributes. That is the right move. If one developer can launch several background sessions, platform teams need to answer basic questions: which parent task spawned this request, which subagent burned the tokens, which workflow got stuck, and which branch or PR did it touch?
This is where Agent View becomes an enterprise signal. Parallelism multiplies productivity, but it also multiplies ambiguity. A single Claude Code session that makes a bad edit is debuggable. Six background sessions editing adjacent parts of a monorepo while one is blocked on an MCP permission and another is pushing a PR is a governance problem. Agent IDs in headers and telemetry do not solve governance by themselves, but they create the data model teams need before writing policy.
Even the row summaries have an operational lesson. Anthropic says each row’s one-line activity summary is generated by a configured Haiku-class model and billed through the normal provider. That is probably cheap, and it is probably the correct UX tradeoff. But it also means the control plane now consumes model calls. Teams modeling Claude Code spend should stop pretending only “code generation” costs money. The supervisor summarizes, classifies, retries, observes, and routes work. Compute is becoming interface.
The plugin and hook updates sit in the same category. claude plugin details <name> now reports component inventory and projected per-session token cost. Hooks can use an args: string[] exec form so commands spawn without a shell. PostToolUse hooks can set continueOnBlock. Hooks now run without terminal access after a bug where hook terminal output could corrupt prompts. These sound like changelog miscellany until you remember what hooks are: lifecycle code around an agent that can run commands, enforce policy, inspect tool use, or alter workflow behavior.
The shell-free hook exec form is particularly important. Quoting bugs are where “helpful automation” becomes command injection. Giving hook authors a way to spawn commands without shell parsing is exactly the kind of unglamorous API hardening that mature agent tooling needs. It will not trend on Hacker News. It will prevent somebody’s policy hook from becoming the exploit path.
MCP reconnects are the plumbing users notice only when they fail
v2.1.139 also tightens MCP behavior in ways that matter for teams treating Claude Code as real tooling. Stdio MCP servers now receive CLAUDE_PROJECT_DIR. Plugin configs can reference ${CLAUDE_PROJECT_DIR}. The /mcp Reconnect command picks up .mcp.json edits without requiring a restart. Remote MCP reconnect retry on transient failure is enabled for all users. Disconnected-server resources stop lingering in @server: autocomplete.
These are not launch-demo features. They are the little cuts that make agent workflows feel unreliable when they go wrong. A stale MCP resource in autocomplete wastes attention. A reconnect that ignores config edits teaches users to restart the whole tool. A missing project directory forces brittle wrappers. If Claude Code is becoming a supervisor for parallel jobs, MCP needs to behave like infrastructure, not a loose collection of adapters.
The security angle is unavoidable. This week’s broader agent-security story is full of repo-controlled MCP config, hook risks, and trust prompts doing too much work. Agent View does not reduce that risk by itself. In some ways, it increases the need for discipline because it makes parallel delegation easier. The right response is not to avoid the feature; it is to pair it with managed settings, telemetry, plugin review, and rules about which repositories and credentials background sessions may touch.
Practitioners should roll this out like a capability expansion, not a UI preference. Start with a small pilot group. Verify that agent_id and parent_agent_id appear in OpenTelemetry. Decide whether regulated teams should disable Agent View by default. Write down when developers may dispatch parallel sessions and when they must attach before continuing. Add explicit guidance for untrusted branches, production credentials, MCP servers, and monorepos. The feature makes it easier to run more work; governance needs to scale with that.
The clean editorial read is this: Claude Code is crossing from coding assistant to supervised agent runtime. The terminal table is just the artifact users can see. Underneath it are queues, persistent sessions, supervisor processes, telemetry identifiers, hook hardening, MCP reconnect semantics, and a model-powered control plane. That is what serious agent tooling was always going to become. The surprise is not that Anthropic built it. The surprise is that the chatbot metaphor lasted this long.
Sources: Claude Code v2.1.139 GitHub release, Claude Code Agent View documentation, Claude Code hooks reference, TestingCatalog