Claude Code Agent View Deep Dive

Agent View is a dashboard, released by Anthropic on May 11, 2026 as a Research Preview, that lets you manage many Claude Code sessions from a single screen. Running claude agents opens a list of running, input-needing, and completed sessions where you can dispatch new tasks, reply to prompts, stop runs, and clean up — all in place.

Requirements: Claude Code v2.1.139+ (check with claude --version). All plans are supported — Pro, Max, Team, Enterprise, and direct API keys. You can reach Agent View from the CLI, the desktop app, the VS Code extension, JetBrains IDEs, the web app at claude.ai/code, and the Slack integration. Org admins can disable it via the disableAgentView managed setting or the CLAUDE_CODE_DISABLE_AGENT_VIEW env var.

You may see this feature called "FleetView" in community posts; that's not the official name and is often confused with the unrelated third-party VS Code extension "Agent Fleet." This article uses the official term Agent View throughout.

Each session shows a state icon: Working / Needs input / Idle / Completed / Failed / Stopped. Different glyphs distinguish a live process (✻), a paused one (∙), and a /loop sleep (✢). A Haiku-class model refreshes the one-line summary on each row roughly every 15 seconds (these summary calls are billed at normal rates).

The input bar at the bottom dispatches new sessions in parallel. Hit Space to open the peek panel and reply, answer prompts, or send a Bash command inline (! prefix). Enter or → attaches; ← detaches and lets the session keep running. Alt+1..9 attaches directly to the Nth session. Ctrl+S toggles grouping by state or directory, Ctrl+T pins, Ctrl+X stops or removes.

From the shell, claude --bg "<prompt>" starts a background session directly; inside an interactive session, /bg (/background) demotes the current run. On the first write, the session is automatically moved into a git worktree under .claude/worktrees/<id>, preventing file conflicts between parallel sessions. A supervisor process keeps the session alive independent of your terminal, so closing the shell does not kill it. Completed sessions that go ~1 hour without an attach have their process stopped, and are respawned on next attach.

claude attach <id> / claude logs <id> / claude stop <id> / claude respawn <id> / claude rm <id> are available for scripting and CI integration.

Sessions dispatched from Agent View pick a subagent internally. The publicly documented built-ins:

Custom subagents live in .claude/agents/*.md. Frontmatter supports model, tools, permissionMode, and isolation: worktree. From the Agent View input bar you launch them with <agent-name> <prompt> or @<agent-name>.

1. Parallel research — fire off several investigations, compare summaries in peek
2. Independent tasks at once — bug fix, PR review, log triage, each in its own worktree
3. Long background jobs — claude --bg "investigate flaky test" and walk away
4. /loop scheduling — cron-like recurring sessions, all visible in one list
5. Multi-repo dispatch — open Agent View from a parent directory and target @<repo>
6. Agent Teams — with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, teammates collaborate via SendMessage and a shared task list (v2.1.32+)

At Oflight we pair Agent View with DocDD (document-driven development) and a ticket-driven flow modeled on OpenAI Symphony.

The docs are blunt: "rate limits apply: background sessions consume your subscription usage the same as interactive sessions." Parallelism is a multiplier on your quota.

Developer write-ups have reported 49 subagents over 2.5 hours burning $8,000–$15,000, and an unattended three-day run hitting $47,000 (claudefa.st, howdoiuseai.com). The row summaries inside Agent View are themselves Haiku calls and bill normally.

Cost hygiene basics:
- Write the parallel-session cap into CLAUDE.md (e.g., 4–8 max)
- Route exploration to Explore + Haiku
- Always include an explicit goal, termination condition, and time budget for unattended runs
- Admins should watch daily quota on the org dashboard

- Large refactors split across worktrees in parallel
- Monorepo-wide impact analysis via Explore agents
- Doc auto-generation — README / OpenAPI per service, dispatched in batch
- Parallel PR review — claude --agent code-reviewer --bg "address review on PR 1234" per PR
- Flaky test investigation — the canonical long-running use case in the official docs
- CI/CD integration — pair headless mode with webhooks so issues auto-spawn agents

- Rate limits scale linearly with parallelism
- Local-only execution — background sessions die on sleep/shutdown (claude respawn --all brings them back)
- Worktrees vanish with the session — merge/push before removing
- No subagent nesting — a subagent cannot call another subagent (Plan is the workaround)
- Subagents return a single message — validation is the parent's responsibility
- bypassPermissions / auto must be approved in advance to keep unattended runs safe
- Interactive sessions running in other terminals do not appear in Agent View
- Subagents themselves are not surfaced as rows in Agent View

Claude Code's edge is the combination of local supervisor + automatic worktree isolation + headless CLI — it composes naturally with CI/CD, cron, and webhooks.

- May 11, 2026 Agent View GA (Research Preview); on the same day major Claude capabilities became available through AWS Marketplace
- Milestones: Agent Teams flag in v2.1.32, git worktree CLI integration in v2.1.50, Agent View in v2.1.139
- VS Code v1.109's Agent Sessions view is converging the multi-agent (Claude/Codex/Copilot) single-pane experience
- Community: 4–8 worktrees per developer reported as a stable line; OSS helpers like johannesjo/parallel-code and spillwavesolutions/parallel-worktrees have appeared
- Anthropic's own engineering blog "Building a C compiler with a team of parallel Claudes" documents their internal practice

Across our AI BPO and AI consulting engagements we treat Agent View as:

1. A human-driven dispatch hub — direct CLI use, standardized
2. Guardrailed unattended runs — worktree + explicit termination prompts to avoid bill shocks
3. Ticket-driven flow — modeled on OpenAI Symphony, implemented with /loop × webhooks
4. Dual-layer cost monitoring — API keys + Anthropic admin dashboard
5. Client deliverable — CLAUDE.md must enumerate parallel-cap, model choice, and termination conditions before handoff

Q1. Is Agent View free?
A. You need a Pro+ subscription or an API key. No extra charge for the feature itself, but every session consumes tokens.

Q2. How many sessions in parallel is safe?
A. No official cap. Community practice reports 4–8 worktrees as a sustainable line balancing cost and review load.

Q3. My background session disappeared.
A. The process stops after ~1 hour without an attach (the session record stays). Use claude respawn <id> or claude respawn --all.

Q4. Where are worktrees stored?
A. Under .claude/worktrees/<session-id>. Removing the session deletes the worktree, so merge/push/branch out first.

Q5. How do I prevent a runaway bill?
A. Four basics: (1) parallel cap and termination conditions in CLAUDE.md, (2) avoid bypassPermissions, (3) switch to API-key billing with daily-budget alerts, (4) route exploration to Haiku/Explore.

Q6. Is FleetView the same thing?
A. The official name is Agent View. "FleetView" is community slang, sometimes confused with the unrelated third-party VS Code extension "Agent Fleet."

Q7. Do my existing Claude Code sessions show up automatically?
A. No. Interactive sessions running in other terminals are not listed. Only sessions started via Agent View, --bg, or /bg appear.

Agent View is the moment Claude Code officially graduated from "one careful session" to "a cockpit for many at once." Automatic worktree isolation, a headless CLI, and Slack/Web dispatch combine to push AI-assisted coding into routine parallel operation. But parallelism is the same multiplier on your bill, and unattended sessions with relaxed permissions can produce four- and five-figure incidents. The new required skill set for teams running Claude Code in 2026 is writing down — and enforcing — your parallel cap, termination conditions, and cost monitoring.

Primary (official):
- Anthropic, "Agent View in Claude Code"
- Claude Code Docs, "Agent View"
- Claude Code Docs, "Sub-agents"
- Claude Code Docs, "Agent Teams"
- Claude Code Docs, "Costs"
- Anthropic Engineering, "Building a C compiler with a team of parallel Claudes"

Secondary:
- TestingCatalog on Agent View
- Verdent, Claude Code Pricing 2026
- claudefa.st, Usage Optimization
- VS Code Blog, Multi-Agent Development
- GIGAZINE, Claude Code Agent View on AWS

Note: the name "FleetView," the Max-plan token estimates (88,000 / 220,000 per 5 hours), and the cost-incident figures ($8,000 – $47,000) come from third-party sources rather than Anthropic's published docs.