How to Manage 10 Claude Code Agents Without Losing Your Mind

The Real Problem Is Not the Number of Agents but the Chaos

Artem Zhutov explains that opening many Claude Code agents in separate browser tabs quickly leads to loss of context, stale sessions, and mental overload. The core issues include unknown activity per tab, constant window switching, loss of state, and the brain handling what should be system‑managed state.

Claude, M1, M2, M3, M4, M5. Which one was doing what?

Root Cause: Tabs Are Not Work Units – Workspaces Are

Tabs lack semantic meaning and stable organization. When agents multiply, tab names become ambiguous, task ownership unclear, status invisible, context unstable, and scheduling impossible.

Layer 1 – cmux: Isolated Workspaces, Not Tabs

cmux is a terminal workspace manager that creates named, isolated workspaces instead of anonymous tabs. Each workspace has its own terminal and can spawn sub‑terminals, all reachable via keyboard shortcuts.

It's not a new tab. It's an isolated workspace.

Named workspaces (e.g., orchestrator, feature‑research, video‑script, daily‑review) allow tasks to be bound to identifiable spaces.

Layer 2 – cmux Is Programmable

Beyond manual switching, cmux exposes system‑level commands that agents can invoke:

cmux list-workspaces
cmux read-screen --workspace workspace:1
cmux send --workspace workspace:1 "what was our progress today"

These commands enable:

Listing current workspaces (named, not numbered tabs).

Reading the screen of a workspace without interrupting the agent.

Sending a message to a workspace, prompting the agent to continue work.

Three commands: list, read, send. That's it.

Layer 3 – The Orchestrator: One Agent Manages the Rest

A single Claude Code agent acts as an orchestrator, launching, managing, and coordinating the other agents. This replaces manual window handling, progress checking, and task assignment.

I talk to one agent. That agent manages the rest.

The orchestrator reads daily notes, spawns sessions, creates workspaces, and forwards commands, reducing human overhead.

Layer 4 – Dashboard: Global Visibility

Using Obsidian and Obsidian Bases, each workspace generates a session file linked to a dashboard that automatically aggregates status, date, title, and related workspace.

Real @obsdmd Bases dashboard – sessions grouped by status: blocked, done, in‑progress, review

The dashboard answers the question, “What are my agents doing now?” by showing status, date, title, and related workspace for each session.

Layer 5 – The Full Loop

The end‑to‑end workflow closes the loop from planning to verification:

Start with a daily note to define today’s tasks.

The orchestrator reads the plan and understands intent.

Sessions are spawned from the plan.

The orchestrator launches corresponding workspaces.

Human reviews progress and results.

If issues arise, comments are added.

The orchestrator relays feedback to the appropriate agent.

Nothing is "done" until you verify it.

This loop turns tasks into continuously tracked objects, embeds the definition of done into the system, and prevents outputs from remaining unverified.

From Tabs to Workspaces – A New Way of Working with AI

Replacing tab‑based interaction with named workspaces, an orchestrator, and an Obsidian dashboard creates a scalable, controllable personal AI operating system. The shift is not a mere tool tip but a systematic upgrade that enables stable collaboration, knowledge retention, and reliable execution.

Future efficiency will depend less on prompt engineering and more on building such integrated AI workflow systems.