If you use Claude Code, you've probably wished for more specialized help — one agent that really understands Kubernetes, another that's a security expert, another that knows modern Python async patterns inside-out. Without bloating your context with 50,000 extra tokens.
That problem is now solved.
KEY RELEASE: wshobson/agents is a massive, production-ready plugin ecosystem for Claude Code delivering 112 specialized AI agents, 72 focused plugins, 146 modular skills, 16 multi-agent orchestrators, and 79 development tools — all designed for minimal token usage and maximum composability.
Why This Is Different
Most agent setups fall into one of two traps: a single monolithic agent that's slow, expensive, and forgets what it's good at — or a messy collection of random tools fighting each other for context space. wshobson/agents takes a fundamentally different approach: granular plugins. You only install what you actually need.
Example: /plugin install python-development loads just 3 Python-specific agents plus 16 relevant skills — roughly 1,000 tokens total. Nothing else. No bloat, no wasted context, no cost overhead.
It also ships two features that change how you work with Claude Code. Agent Teams run 4–7 specialized agents in parallel (review, debug, feature, full-stack, security). Conductor turns Claude into a structured project manager following a strict Context → Spec & Plan → Implement → Verify pipeline. Both push Claude Code from single-prompt tool to coordinated engineering workflow.
The Three-Tier Model Strategy
The repo intelligently assigns the right model to the right job. 42 agents run on Opus 4.6 for critical architecture and security decisions. 51 agents run on Sonnet 4.6 for balanced everyday work. 18 agents run on Haiku 4.5 for fast operational tasks. The rest inherit whatever model you're already running.
This tiered approach mirrors exactly what NVIDIA's Orchestrator-8B research proved: a trained router that assigns tasks to the right model beats a single frontier model at 30% of the cost. wshobson/agents puts that principle into practice for Claude Code users today.
Install what you need. Nothing else. One thousand tokens per plugin.
What This Means for TRAgents
At TrainingRun.AI, we're already testing several of these plugins in our own agent workflows. This ecosystem directly feeds two TRAgents pillars: Multi-Model Efficiency (the tiered Opus/Sonnet/Haiku routing) and Tool Reliability (146 modular skills with progressive disclosure). Expect relevant metrics in upcoming TRAgents updates.
Get Started
Two commands:
/plugin marketplace add wshobson/agents
/plugin install python-development javascript-typescript full-stack-orchestrationGitHub (29.9K stars): github.com/wshobson/agents
Check out the project: wshobson/agents
View on GitHub →What do you think? Drop a reply on X. We read every one.
— The TrainingRun.AI Team
