Claude Code subagents are specialized sub-agents defined in .claude/agents/*.md that run in their own isolated context and delegate narrow, well-scoped tasks. AgentsRoom is the visual command center, the multi-agent dashboard and the subagent manager that lets you run dozens of Claude Code subagents in parallel, across every project, with zero terminal juggling.
Subagents are Anthropic's answer to context bloat and role specialization. AgentsRoom is the visual layer on top: instead of watching subagents scroll past in a single terminal, you get a real-time dashboard where every Claude Code subagent, every session, every project is visible at a glance. Subagents for focus, AgentsRoom for orchestration.
Claude Code subagents (sometimes written as sub-agents) are specialized AI agents that a parent Claude Code session can delegate work to. Each subagent is defined in a Markdown file inside .claude/agents/ (project-level) or ~/.claude/agents/ (user-level), with its own name, description, system prompt, and allowed tools. When the parent agent needs a narrow task done — a code review, a security audit, a refactor, a test run — it calls the appropriate subagent, which runs in an isolated context window and returns a result.
The key design decision is context isolation. A subagent starts with a clean slate, focused entirely on its narrow job. It doesn't pollute the parent's context with exploration noise, and it doesn't lose its focus to unrelated tangents. This is why Claude Code subagents are the recommended way to scale Claude Code beyond a single long-running conversation: you split big problems into role-based specialized sub-agents and let Claude Code orchestrate them.
Subagents also enable true multi-agent workflows. A single Claude Code session can fan out to multiple subagents, each running a slice of the job, each with its own tools, its own context, its own specialization. This is Anthropic's native primitive for multi-agent coding, and AgentsRoom is the visual command center that takes it to the next level by letting you run many Claude Code sessions (each with their own subagent configs) in parallel, at fleet scale.
Subagents can live at the project level, the user level, or be orchestrated across sessions by a tool like AgentsRoom.
Drop a Markdown file in .claude/agents/ inside your repository and every Claude Code session that runs against that project automatically gains access to your custom subagents. Because the folder is committed to git, your whole team inherits the same specialized sub-agents. This is the right layer for project-specific subagents: a reviewer that knows your conventions, a tester that knows your test framework, a security auditor that knows your threat model.
Subagents defined in your user home directory are available across every project, regardless of which repo Claude Code is running in. This is the right layer for personal productivity subagents: a prompt-formatter, a commit-message-writer, a code-explainer, a TypeScript-strictness enforcer, or any other generalist helper you want in every session.
AgentsRoom doesn't replace .claude/agents subagents — it runs the real Claude Code CLI, so every subagent file you've defined still works exactly the same. What AgentsRoom adds is a third layer: fleet-level orchestration. You can run many Claude Code sessions in parallel, each with their own subagents, each assigned to a different role from AgentsRoom's 14 built-in agent roles (Frontend, Backend, QA, DevOps, Security, Architect…), across multiple projects at once. It's subagents inside sessions, sessions inside the multi-agent dashboard.
Subagents are the most important Claude Code feature for teams that want to scale AI coding without drowning in context.
Each Claude Code subagent starts with a clean, minimal context dedicated to its narrow job. No irrelevant exploration, no stale context, no lost focus. Subagents stay on task because the task is literally all they know.
A parent Claude Code session can dispatch multiple subagents that run independently. Combined with AgentsRoom's ability to launch many Claude Code sessions in parallel, you get true multi-agent, multi-subagent coding: dozens of specialized AI workers, each on their own slice of the problem.
Because subagents are just Markdown files in .claude/agents, they're trivially reusable, versioned with git, and shareable with your entire team. Write the subagent once, commit it, and every developer on the project inherits the same specialized sub-agent. This is multi-agent AI infrastructure as code.
Big codebases are impossible to fit in a single context window. Claude Code subagents solve this by splitting work into narrow-context specialists. AgentsRoom scales the approach further: big codebases become many Claude Code sessions, each running subagent-powered workflows, each visible on the multi-agent dashboard.
Without a visual layer, subagents can feel like a black box: they spin up, do something, and return a result. AgentsRoom surfaces everything Claude Code streams — including SubagentStop events, tool calls, and status changes — in real time, per agent, per project. Subagent execution becomes fully observable.
Claude Code subagents compose beautifully with Claude Code hooks (PreToolUse, PostToolUse, SubagentStop) and with MCP servers. A subagent can respect project-wide hooks, use project-wide MCP servers, and run inside an AgentsRoom-managed session — all without any special configuration on your end. This is Anthropic's native stack, fully respected.
AgentsRoom is the visual command center for Claude Code. It spawns real claude CLI processes inside full xterm.js terminals, which means every subagent you've defined in .claude/agents/*.md or ~/.claude/agents/*.md is automatically available, exactly as Anthropic designed it. No re-implementation, no parsing, no loss of fidelity.
On top of native subagent support, AgentsRoom gives you fleet-level orchestration: 14 built-in agent roles you can assign to sessions, a visual multi-agent dashboard, per-agent notifications, a mobile companion, a drag-and-drop project manager, and a backlog task board that spawns Claude Code agents from kanban cards. It's the only way to run and watch a small army of subagent-powered Claude Code agents without losing your mind.
Every AgentsRoom agent is a real claude process. Your .claude/agents/*.md subagents, your user-level ~/.claude/agents/*.md subagents, your hooks, your CLAUDE.md, your settings — all untouched, all respected, all running as Anthropic intended.
Assign each AgentsRoom session to one of 14 specialized roles: DevOps, Fullstack, Frontend, Backend, Architect, QA, Marketing, PM, Security, Mobile, Git, SEO, i18n, or Custom. Role prompts compose cleanly with your in-session Claude Code subagents for two layers of specialization.
AgentsRoom runs many Claude Code sessions in parallel as separate processes. Each session can dispatch its own subagents. This is real parallelism, not simulated concurrency: your multi-agent Claude Code workflow scales to whatever your machine and API quota allow.
Jump between projects and instantly see which subagents are running, which are blocked on input, which have just finished. The sidebar aggregates status across every project; the dashboard surfaces per-agent activity. Subagents become visible, measurable, manageable.
Download AgentsRoom and run a fleet of Claude Code subagents across all your projects with one visual dashboard. Native .claude/agents support, zero configuration.
Companion app: monitor your agents on the go
Works with Claude, Codex, OpenCode, Gemini CLI, and Aider