Claude Code Agents: What They Do and Which Plugins Use Them

What Are Claude Code Agents?

Agents are one of Claude Code's nine component types — and with 49,907 indexed across the directory, they're the second-largest category after skills (178,848). An agent is a subagent definition: a packaged prompt-and-tooling configuration that Claude Code can spawn to handle a focused task. Unlike skills (which extend Claude's knowledge) or commands (which provide slash-command interfaces), agents run as autonomous sub-processes that can read files, execute code, and return results to the parent session.

The agents directory currently catalogs these components from across 33,537 plugins. Their growth reflects a shift in how developers use Claude Code — from single-session conversations to orchestrated multi-agent workflows where specialized subagents handle discrete responsibilities.

How Agents Work in Claude Code

Claude Code's agent system follows a delegation model. When a plugin defines an agent, it specifies:

• A focused task scope — what the agent is responsible for (e.g., "locate relevant code," "review a diff," "run tests")
• Tool access — which tools the agent can use (file reads, bash commands, MCP servers)
• Prompt context — instructions that shape the agent's behavior and output format

At runtime, Claude Code spawns the agent as a subagent with its own context window. The parent session delegates work, the agent executes independently, and its results flow back. This is how plugins like caveman achieve 75% token reduction — by routing tasks to tightly-scoped subagents instead of handling everything in a single, context-heavy session.

Agents frequently combine with other component types. A plugin might define commands that trigger agents, skills that inform agent behavior, or hooks that fire agents at specific lifecycle events. Of the 304,608 total components in the directory, agents make up roughly 16% — a substantial share that reflects their utility as building blocks for complex workflows.

Agent Plugins Worth Exploring

Here are notable plugins that demonstrate different approaches to using Claude Code agents. Each is selected from the directory's data to illustrate the range of tasks agents can address.

Token-Efficient Coding with caveman

caveman (923 installs in the past 7 days, 51,378 stars) takes a distinctive approach: it switches Claude Code into a terse "caveman mode" that targets 75% token reduction. It ships agents, skills, and hooks as component types. The agents handle code location, surgical 1–2 file edits, and focused diff/PR reviews — each scoped narrowly enough to avoid the context bloat that comes from doing everything in one session. It also auto-generates conventional commits and provides real token usage stats.

Full-Stack Workflow Orchestration with everything-claude-code

everything-claude-code (425 installs/7d, 176,491 stars) is the broadest agent plugin in the directory. It provides commands, agents, skills, hooks, and MCP server components — all five of the primary component types. With 300+ components spanning autonomous multi-agent coding workflows, TDD enforcement, security audits, and production code generation across JS/TS/Python/Rust/mobile stacks, it demonstrates how agents can serve as orchestration points for complex development pipelines.

Spec-Driven Development with gsd-build

gsd-build (101 installs/7d, 57,510 stars) automates spec-driven software development through commands and agents. Its agents autonomously plan, execute, and audit development phases — bootstrapping projects with AI-generated roadmaps, managing git workflows, and generating tests and documentation. The plugin provides 100+ commands and agents, making it one of the more comprehensive workflow-automation tools in the directory.

Documentation Retrieval with context7-plugin

context7-plugin (51,792 stars) uses agents alongside commands, skills, and an MCP server to fetch up-to-date, version-specific documentation for libraries like React, Next.js, Vue, Prisma, and Supabase. Its agents pull API references and code examples directly into your LLM context — solving the stale-training-data problem that affects any model working with fast-moving frameworks.

Prompt Management with prompts.chat

prompts.chat (159,878 stars) exposes commands, agents, skills, and an MCP server for searching, retrieving, and managing thousands of AI prompts. Its agents handle prompt discovery and enhancement — you can install skills to extend capabilities, fill prompt variables, save custom prompts with metadata, and use AI to improve existing prompts.

Plugin Development with plugin-dev

plugin-dev (83,183 stars) helps you build Claude Code plugins themselves. It provides commands, agents, and skills for implementing agents, commands, skills, and hooks with MCP integration. Its agents run creation workflows, generate validation reports on plugin structure, and review skills for adherence to conventions — a meta-tool for the plugin ecosystem.

Emerging: Dedicated Agent Infrastructure

Two newer plugins focus on agent infrastructure specifically. agent-orchestration provides agents and skills for coordinating multi-agent workflows, while agent-cost offers agents and skills for tracking the cost of agent execution. Both are early-stage (0 stars each), but they signal growing interest in tooling that manages agents as first-class resources rather than embedded features of larger plugins.

Browse the full list of agent-enabled plugins in the plugins directory or filter by component type on the agents page.

Patterns to Borrow from Agent Plugins

Looking across these plugins, several agent design patterns emerge that are worth adopting in your own work:

Narrow scope, clear handoff. The most effective agent plugins give each agent a single responsibility. caveman doesn't have one agent that does everything — it has separate agents for code location, editing, and review. This keeps context windows small and outputs predictable.

Combine agents with other component types. Five of the six established plugins above ship agents alongside at least two other component types. everything-claude-code uses all five primary types (commands, agents, skills, hooks, MCP). Agents work as the execution layer, but commands give users entry points, skills provide domain knowledge, hooks automate triggers, and MCP servers connect to external data.

Use agents for the expensive parts. caveman's 75% token reduction comes from moving work out of the main session. If a task requires reading many files or processing large outputs, spawning a subagent to handle it is cheaper than expanding the parent context.

Let agents compose. gsd-build's 100+ commands and agents don't work in isolation — they chain together for spec-to-plan-to-build-to-test-to-release workflows. Designing agents that produce structured output lets downstream agents consume their results without re-parsing.

Where Claude Code Agents Are Headed

With 49,907 agents across the directory and 1,977 new plugins added this week alone, the agent component type is well-established in the Claude Code ecosystem. The emergence of dedicated infrastructure plugins like agent-orchestration and agent-cost suggests the ecosystem is maturing past "agents as features" toward "agents as managed resources" with their own coordination and cost layers.

For developers evaluating agent plugins to install: start with your workflow bottleneck. If token costs are the pain point, caveman's delegation model is worth testing. If you need breadth across languages and frameworks, everything-claude-code covers the widest ground. If you want structured, spec-driven development, gsd-build's autonomous planning agents offer a clear workflow. And if you're building plugins yourself, plugin-dev's agent-powered scaffolding and validation can accelerate development.

The agents page lets you browse all 49,907 agents by plugin, and each plugin's detail page shows its full component breakdown.