Agent View & Dynamic Workflows

Two ways to scale beyond one conversation: agent view manages many background sessions from one screen; dynamic workflows orchestrate dozens-to-hundreds of subagents from a script.

Agent View (claude agents)

One screen for all your background sessions: what’s running, what needs input, what’s done. Each row is a full Claude Code session that keeps running with no terminal attached (a supervisor process hosts them — they survive closing your shell and machine sleep).

claude agents                    # open agent view
claude agents --cwd ~/proj/app   # only sessions from one project

Key	Action
type + Enter	Dispatch a new background session (each prompt = new session)
↑/↓ + Space	Peek: latest output / pending question; type to reply, number keys answer multiple-choice
Enter / →	Attach (full interactive session; recap of what you missed)
← on empty prompt	Detach back to the table (or background the current session)
Ctrl+X	Stop the selected session

Inside any normal session, /bg (or ←) backgrounds it into agent view. Rows show state (working / needs input / completed / failed), a Haiku-generated one-line summary, and PR status labels (#2048 colored yellow = checks pending, green = ready, purple = merged) — for most tasks the green PR number is where you pick up the result.

Agent view as your main entry point: dispatch every task from the table, attach only when a session needs real conversation.

Dynamic Workflows

A workflow is a JavaScript script that orchestrates subagents — Claude writes it, a runtime executes it in the background, and your context only receives the final result. For work bigger than any single conversation: codebase-wide audits, 500-file migrations, cross-checked research.

v2.1.154+, all paid plans (Pro: enable in /config).

Who holds the plan?

	Subagents	Agent teams	Workflows
Orchestrator	Claude, turn by turn	Lead agent	The script
Intermediate results	Claude’s context	Shared task list	Script variables
Scale	A few per turn	3–5 peers	Dozens–hundreds per run

Triggering

/deep-research What changed in the Node.js permission model between v20 and v22?

(bundled workflow: fans out searches, cross-checks sources, returns a cited report)

ultracode: audit every API endpoint under src/routes/ for missing auth checks

(the ultracode keyword — or just say “use a workflow” — makes Claude write one for your task)

/effort ultracode

(xhigh reasoning + Claude auto-plans workflows for every substantive task — expensive, powerful)

Managing runs

• /workflows → progress view: phases, agent counts, token totals; p pause/resume, x stop, r restart agent, s save as a reusable command (.claude/workflows/ project or ~/.claude/workflows/ personal)
• Saved workflows run as /<name> and accept args: Run /triage-issues on issues 1024, 1025, and 1030
• Limits: max 16 concurrent agents, 1,000 agents/run, no mid-run user input; workflow subagents always run in acceptEdits with your tool allowlist
• Cost scales with agents — pilot on a small slice first

Example prompts that fit workflows

> use a workflow to run npx tsc --noEmit and keep fixing errors until it
  passes or two rounds in a row make no progress
> use a workflow to migrate every component under src/components/ from
  styled-components to Tailwind, each file in its own isolated copy
> use a workflow to review every file changed in this PR, then merge the
  per-file findings into one ranked summary

Which Parallelism Tool When?

Need	Use
A focused helper that reports back	Subagent (29)
Peers that debate/coordinate	Agent team (32)
Many independent full sessions, managed from one screen	Agent view
Repeatable orchestration of 10s–100s of agents	Workflow
Manual parallel sessions on branches	Worktrees (31)

🧠 Quick Recall

• Trick: “Agent view = mission control; a workflow = the script holds the plan.”

👨‍🏫 Teach It

• Show: claude agents → dispatch 3 tasks → peek with Space.
• They’ll trip on: every prompt in agent view starts a NEW session — peek/attach to follow up.

---

Learning path: ⬅ 49-goal-and-advisor.md · Index · ➡ 51-artifacts-and-deep-links.md