@agent-smith/feat-agents

Multi-agent coordination system for the Agent Smith toolkit — 16 specialized AI agents, 4 workflows, and 15+ reusable skills for task management and project automation.

This package is part of the Agent Smith project, a TypeScript toolkit for running and managing AI agents.

📖 Documentation

For AI Agents

• Codebase Summary — Architecture, agent definitions, skills, and patterns for the agents plugin
• Project Navigation — Map of all Agent Smith repositories and plugins

For Humans

• Plugin Overview — Complete plugin documentation
• GitHub Repository — Source code and contributions

✨ Features

• 🤖 Coordinator Agent — Central agent-smith agent that decomposes tasks, plans workflows, and delegates to specialized agents via tool calls
• 🔍 16 Specialized Agents — Pre-configured agents for search, code, documentation, SQL queries, assistance, help, state tracking, project management, and collaboration
• ⚡ 4 Workflows — YAML-defined pipelines for config info retrieval, database queries, Q&A sessions, and vision tasks
• 📋 15+ Reusable Skills — Task creation/execution (solo/team), documentation updates, codebase summaries, and project management
• 🧩 Context Fragments — Workspace context helpers that provide agents with project structure awareness
• 🔗 Agent Delegation — Coordinator calls specialized agents sequentially via the run-agent tool
• 🎯 Model Flexibility — Supports all compatible LLM models (qwen35b, qwen4b, etc.) with configurable inference parameters
• 🛡️ Security Patterns — Agents operate within a specified workspace directory with controlled tool access

📦 Installation

# Install globally via npm
npm i -g @agent-smith/feat-agents

Configuration

Add the plugin to your config.yml:

plugins:
  - "@agent-smith/feat-agents"

Then update your client configuration:

lm conf

🚀 Quick Start

Launch the main coordinator agent:

lm agent-smith "Create documentation for my project"

The coordinator will automatically decompose your request and delegate to specialized agents as needed.

For direct inference (bypassing coordination):

lm infer "Explain how multi-agent coordination works"

📋 Agents Reference

Core Agents

| Agent | Description | |-------|-------------| | agent-smith | Main coordinator — Orchestrates team of agents, decomposes tasks, delegates work | | agent-smith-assistant | Default assistant agent — Fork of coordinator for general task execution | | agent-smith-light | Lightweight agent for simpler tasks with reduced context | | infer | Raw inference agent for direct LLM queries without orchestration |

Specialized Agents

| Agent | Description | |-------|-------------| | agent-smith-search | Web search and webpage reading — uses ddsearch, read-webpage tools | | agent-smith-code | Code management agent for programming tasks | | agent-smith-doc | Documentation management agent for project docs | | agent-smith-sql | Database query agent — generates SQL from natural language questions | | agent-smith-help | Help agent with full capabilities | | agent-smith-help-light | Lightweight help agent | | agent-smith-help-assistant | Delegated work assistant for help scenarios | | agent-smith-state | Project state management — checks context and documentation completeness | | agent-smith-project | Project management agent | | agent-smith-colab | Collaborative agent (experimental) | | agent-smith-doc-colab | Documentation collaboration agent | | collaborator | Worker collaborator agent for role-based tasks |

🔄 Workflows

The plugin includes 4 workflow definitions for multi-step operations:

| Workflow | Purpose | |----------|---------| | agent-smith-config-info | Configuration information retrieval workflows | | agent-smith-db | Database query orchestration with SQL agent | | q | Q&A workflow — adapts pre-query then runs inference agent | | vision | Vision tasks workflow for image/video analysis |

Execute a workflow:

lm workflow agent-smith-config-info

🛠️ Skills Reference

The plugin provides 15+ reusable skills for AI coding agents, loaded via the load-skill tool.

Task Creation

| Skill | Description | |-------|-------------| | create-task | Entry point — analyzes request, chooses execution mode, creates task | | create-task-solo | Single-agent task creation | | create-task-team | Multi-agent team task creation | | create-task-from-template | Create tasks from predefined templates | | use-tasks | Task usage and management skill |

Task Execution

| Skill | Description | |-------|-------------| | execute-task | Task execution (uses phase-specific skills) | | execute-task-solo | Single-agent task execution | | execute-task-team | Multi-agent team task execution | | execute-task-phase | Phase-specific execution logic | | task-mode | Decision skill — evaluates task complexity to choose solo vs team mode |

Documentation

| Skill | Description | |-------|-------------| | create-readme | Create comprehensive README files for packages/plugins/repositories | | document-package | Create or update package documentation in the docsite | | create-project-docs | Generate project documentation structure | | update-codebase-summary | Create or update codebase summary documentation | | update-project-nav | Update the project navigation map |

Exploration

| Skill | Description | |-------|-------------| | smart-explore | Navigate and explore module directory trees efficiently |

📋 Tasks Reference

The plugin includes pre-defined tasks — structured work units with goals, plans, and progress tracking for complex multi-step operations. Use the load-task tool to load a task into your workspace, then execute it with the appropriate skill.

| Task | Description | Mode | Phases | |------|-------------|------|--------| | update-project-doc | Create or update AI agent navigation documentation for a project (AGENTS.md, decision-tree.md, codebase-summary.md, project-nav.md, etc.) | Solo | 8 phases: Explore → AGENTS.md → Per-Repo AGENTS.md → decision-tree → project-overview → codebase-summary → project-nav → Cross-Reference |

Using Tasks

Prompt to execute a task:

execute the update-project-doc

%execute-task-solo

The %execute-task-solo part loads this skill directly in the prompt, so that the agent has all the execution instructions. Use %execute-task-team for multi agents tasks. Note: the agent must have the load-task tool. When it uses it the task directory is copied to .agents/tasks/[task-name] and it's execution and state will be managed from there.

The agent will execute phases sequentially, verify success criteria after each step, and update state.md to track progress.

Task Structure

Each task contains:

• goals.md — Goals and success criteria
• plan.md — Step-by-step execution plan with verifiable success criteria
• state.md — Progress tracking (updated after each phase)
• notes.md — Optional context, rules, and related skills

📁 Structure

agents/
├── package.json                    # Plugin metadata (@agent-smith/feat-agents)
└── dist/
    ├── agents/                     # 16 agent YAML definitions
    │   ├── agent-smith.yml         # Main coordinator agent
    │   ├── agent-smith-assistant.yml
    │   ├── agent-smith-search.yml
    │   ├── agent-smith-code.yml
    │   ├── agent-smith-doc.yml
    │   ├── agent-smith-sql.yml
    │   ├── agent-smith-help.yml
    │   └── ... (16 total)
    ├── workflows/                  # 4 workflow definitions
    │   ├── agent-smith-config-info.yml
    │   ├── agent-smith-db.yml
    │   ├── q.yml
    │   └── vision.yml
    ├── skills/                     # 15+ skill modules
    │   ├── create-readme/SKILL.md
    │   ├── create-task/SKILL.md
    │   ├── smart-explore/SKILL.md
    │   └── ... (16 total)
    ├── tasks/                      # Pre-defined task definitions
    │   └── update-project-doc/     # Project documentation task
    │       ├── goals.md
    │       ├── plan.md
    │       ├── state.md
    │       └── notes.md
    └── fragments/                  # Context helper files
        ├── workspace.md            # Workspace context info
        └── ctx-helper-files.md     # Context file references

🔧 Architecture

Agent Coordination Pattern

The agent-smith coordinator agent follows a structured workflow:

1. Analyze the user request
2. Decompose into subtasks if necessary
3. Plan the work sequence
4. Delegate to specialized agents via run-agent tool (one at a time)
5. Synthesize results from agent responses

Available Tools

Agents have access to these tools:

| Tool | Description | |------|-------------| | readfile | Read file contents | | edit-search-replace | Search and replace in files | | shell | Execute shell commands | | python | Execute Python code | | run-agent | Call other agents for delegated work | | load-skill | Load skill modules for specific tasks | | load-task | Load task definitions for complex operations | | notify-user | Send notifications to the user |

Context Fragments

Agents can include project context using file references in YAML:

template:
    system: |-
      {file:../fragments/workspace.md}
      {file:../fragments/ctx-helper-files.md}

Sample Agent Definition

Each agent is defined as a YAML file:

name: infer
description: Run a raw inference query
type: "agent-smith"
prompt: |-
    {prompt}
model: qwen4b
inferParams:
  top_p: 0.95
  top_k: 20
  min_p: 0
  temperature: 0.6
  repeat_penalty: 1

⚙️ Configuration

Model Parameters

Agents support configurable inference parameters:

inferParams:
    min_p: 0
    top_k: 20
    top_p: 0.85
    temperature: 0.6
    repetition_penalty: 1
    presence_penalty: 1.5
chat_template_kwargs:
    enable_thinking: true
    preserve_thinking: true

Variables

Agents can require runtime variables:

variables:
    required:
        workspace:
            description: The local directory path where to operate

Customizing Agent Behavior

To customize an agent, copy the YAML from dist/agents/ to your local features directory and modify it. The coordinator uses a system prompt that lists available agents and their purposes:

template:
    system: |-
      You are Agent Smith, an AI coordinator agent. Your job is to organise
      the workflow of other agents. Available agents:
      - `agent-smith-assistant`: default assistant
      - `agent-smith-search`: web search and reading
      {file:../fragments/workspace.md}

🔗 Related Packages

| Package | Description | |---------|-------------| | @agent-smith/core | Core framework — workflow engine, agent runtime, tool integration | | @agent-smith/cli | Terminal client that consumes plugins | | @agent-smith/feat-git | Git operations with AI-powered commit messages | | @agent-smith/feat-sqlite | SQLite database operations | | @agent-smith/feat-search | Web search and crawling tools (used by search agent) | | @agent-smith/feat-shell | Sandboxed shell/Python execution in Docker containers | | @agent-smith/feat-fs | Filesystem operations with path authorization |

⚠️ Important Notes

• No runtime code — Agents, workflows, and skills are YAML/Markdown files loaded by the @agent-smith/core framework at runtime
• YAML-based configuration — All agent definitions use YAML for easy customization
• Sequential delegation — The coordinator calls one agent at a time for task completion
• Model flexibility — Supports all compatible LLM models with configurable inference parameters
• Workspace context — Agents operate within a specified workspace directory (required variable)
• Requires core framework — This plugin depends on @agent-smith/core for execution

📄 License

MIT