Multi-Agent Orchestration
by Seth Hobson
Spawn three parallel code reviewers, each dissecting your pull request from a different angle -- security, performance, and architecture -- then get a consolidated report with prioritized findings in under two minutes. That is what Multi-Agent Orchestration from wshobson/agents actually delivers inside Claude Code. This skill is the centerpiece of a massive 72-plugin ecosystem (112 agents, 146 skills, 79 tools) that transforms Claude Code from a single-threaded assistant into a coordinated team of specialists working simultaneously. The core idea is deceptively simple: instead of asking Claude to handle a complex task sequentially, you decompose the work across multiple agents that operate in parallel with strict file ownership boundaries. A team-lead agent handles decomposition and result synthesis. Specialized team-reviewer, team-debugger, and team-implementer agents execute their assigned work concurrently. The result is not just faster execution -- it is qualitatively better output, because each agent operates within a focused domain rather than trying to hold an entire codebase in context at once. Seven preset team configurations cover the workflows developers hit most: multi-reviewer code analysis, hypothesis-driven debugging (where agents compete to find root cause with evidence), parallel feature development with file ownership to prevent merge conflicts, full-stack end-to-end development, coordinated research across codebase and documentation, security audits covering OWASP top 10 plus dependency chains, and structured codebase migrations with verification gates. You can also compose custom teams. Setup requires enabling Claude Code's experimental agent teams feature via an environment variable and choosing a display mode -- tmux panes (recommended, each agent gets its own visible terminal), iTerm2 tabs on macOS, or in-process. The plugin installs through Claude Code's plugin marketplace in two commands. No external MCP servers are required; the entire system is self-contained within Claude Code's plugin architecture. The four-tier model strategy is worth noting: critical architecture and security decisions route to Opus 4.6, complex implementation uses your session's model, support tasks like testing go to Sonnet 4.6, and fast operational work runs on Haiku 4.5. This means you are not burning expensive tokens on tasks that do not need them. With 30.3K GitHub stars and active development (240+ commits), this is the most mature multi-agent orchestration system available for Claude Code in 2026.
Installation
Key Features
- ✓Spawns parallel agent teams with seven presets (review, debug, feature, fullstack, research, security, migration) so you pick a workflow and agents self-organize around it
- ✓File ownership boundaries prevent merge conflicts during parallel feature development -- each team-implementer agent writes only to its assigned files, eliminating the coordination headache of concurrent edits
- ✓Four-tier model routing sends architecture decisions to Opus 4.6, complex tasks to your session model, support work to Sonnet 4.6, and operational tasks to Haiku 4.5, cutting token costs without sacrificing quality where it matters
- ✓Hypothesis-driven debugging assigns competing agents to independently investigate different root cause theories with evidence gathering, then arbitrates findings -- significantly faster than sequential elimination
- ✓Team-lead agent handles task decomposition, dependency graphing, workload balancing, and result synthesis automatically, so you describe the goal and get back a consolidated deliverable
- ✓72 composable plugins (112 agents, 146 skills, 79 tools) across 24 categories including Python, TypeScript, Kubernetes, security scanning, LLM applications, and infrastructure -- install only what you need
- ✓Conductor plugin transforms Claude Code into a project management layer with context-driven track-based development, interactive setup, and semantic revert capabilities for long-running projects
- ✓Live team monitoring via /team-status command shows each agent's progress, assigned tasks, and completion state, with graceful /team-shutdown for clean resource cleanup
Use Cases
- →You have a 2,000-line pull request that needs review before merging -- instead of one pass catching surface issues, spawn a review team where three agents analyze security vulnerabilities, performance bottlenecks, and architectural consistency in parallel, then get a unified prioritized report
- →A production bug has three plausible root causes and you are wasting time testing them one by one -- launch a debug team where agents independently investigate each hypothesis with evidence gathering, and the team-lead arbitrates which theory the evidence supports
- →You need to build an OAuth2 authentication system spanning backend API, database migrations, frontend components, and tests -- use the fullstack preset to decompose the work across specialized agents with strict file ownership, avoiding the merge hell of parallel manual development
- →Your codebase needs migration from Express to Fastify across 40+ route files -- the migration preset plans the transformation, parallelizes file-by-file conversion across agents, and runs verification gates to catch regressions before you commit
- →A security audit is due and you need OWASP coverage, dependency vulnerability scanning, secrets detection, and authentication flow review -- the security team preset assigns each dimension to a dedicated agent and consolidates findings into an actionable remediation plan
- →You are researching whether to adopt a new library and need to evaluate documentation quality, community activity, API surface, and integration complexity simultaneously -- the research preset dispatches investigators to explore each dimension in parallel