agentk8 2.3.3

Multi-LLM Council Backend for AGENT-K - Three-stage consensus with GPT, Gemini, and Claude

AGENT-K

Multi-Agent Claude Code Terminal Suite

╭─────────────────────────────────────────────────────────────────────────────────╮
│                              AGENT-K v1.0                                       │
╰─────────────────────────────────────────────────────────────────────────────────╯

Transform your terminal into a team of specialized AI agents. AGENT-K orchestrates multiple Claude instances working in parallel on your software development and ML research tasks.

What is AGENT-K?

AGENT-K extends Claude Code with multi-agent orchestration. Instead of a single AI assistant, you get a coordinated team of specialists - each with domain expertise, working together on complex tasks.

╭─ You ────────────────────────────────────────────────────────────────────────────╮
│ Build a secure REST API with user authentication                                 │
╰──────────────────────────────────────────────────────────────────────────────────╯

╭─ Orchestrator ───────────────────────────────────────────────────────────────────╮
│ Breaking down task...                                                            │
│   → Engineer: Implement REST endpoints and JWT authentication                    │
│   → Tester: Write API integration tests                                          │
│   → Security: Review for OWASP vulnerabilities                                   │
│   → Scout: Find latest best practices for JWT in 2025                            │
╰──────────────────────────────────────────────────────────────────────────────────╯

[●] Orchestrator   Coordinating
[◐] Engineer       Writing src/api/auth.py...
[◐] Scout          Searching for JWT best practices...
[ ] Tester         Waiting for implementation
[ ] Security       Queued for review

Why AGENT-K?

The Problem

Claude is brilliant, but complex tasks often require:

• Multiple perspectives (implementation, testing, security)
• Parallel work (why write tests sequentially after code?)
• Up-to-date information (Claude's training data becomes stale)
• Specialized focus (security reviews need different prompts than coding)

The Solution

AGENT-K spawns specialized Claude agents that:

• Work in parallel on different aspects of your task
• Coordinate automatically through a central orchestrator
• Stay current with a dedicated Scout agent for real-time research
• Follow best practices with domain-specific system prompts

Features

Feature	Description
Multi-Agent Orchestration	Coordinate 5-6 specialized agents working in parallel
Two Modes	Software Development (default) and ML Research & Training
Interactive Chat	Familiar interface like claude but with a whole team
Visual Mode	tmux-based multi-pane view of all agents working
Scout Agent	Real-time web/GitHub/paper search to stay current
Date Awareness	Agents know when to verify potentially outdated info
Focus Mode	Talk directly to any specialist agent
File-Based IPC	Agents coordinate through structured JSON messages

Agent Teams

Development Mode (Default)

Agent	Specialty
Orchestrator	Task decomposition, coordination, result aggregation
Engineer	Code implementation, debugging, refactoring
Tester	Unit/integration tests, coverage analysis
Security	OWASP vulnerability review, secrets detection
Scout	Real-time search for current best practices

ML Mode (--mode ml)

Agent	Specialty
Orchestrator	ML project lifecycle management
Researcher	Literature review, SOTA analysis, paper summaries
ML Engineer	Model implementation, training loops, optimization
Data Engineer	Data pipelines, preprocessing, augmentation
Evaluator	Metrics, benchmarking, experiment tracking
Scout	arXiv, HuggingFace, Papers With Code search

Installation

Homebrew (macOS/Linux)

brew tap de5truct0/agentk
brew install agentk8

npm

npm install -g agentk8

pip

pip install agentk8

Quick Install Script

curl -sSL https://raw.githubusercontent.com/de5truct0/agentk/main/install.sh | bash

From Source

git clone https://github.com/de5truct0/agentk.git
cd agentk
make install

Note: Package name is agentk8 on all registries. The installed command is agentk.

Requirements

• jq - JSON processing (brew install jq)
• claude - Claude Code CLI (Install here)
• tmux - Optional, for visual mode (brew install tmux)

Quick Start

# Start interactive session
agentk

# Start ML research mode
agentk --mode ml

# Start with visual panels (requires tmux)
agentk --visual

# One-shot task
agentk -c "Refactor the user service to use async/await"

Usage

Interactive Session

$ agentk

╭─────────────────────────────────────────────────╮
│  AGENT-K v1.0                                   │
╰─────────────────────────────────────────────────╯
Mode: Software Development Mode

Type your request or /help for commands.

╭─ You ─────────────────────────────────────────────────
│

Session Commands

Command	Description
/status	Show all agent states and current tasks
/logs <agent>	View agent output
/kill <agent|all>	Stop agent(s)
/focus <agent>	Talk directly to one agent
/unfocus	Return to orchestrator
/visual	Toggle tmux panel view
/clear	Clear screen
/help	Show all commands
/exit	End session

Scout Commands (Both Modes)

Command	Description
/search <query>	Web search for latest info
/github <query>	Search GitHub repos and code
/papers <topic>	Search arXiv/Semantic Scholar
/libs <task>	Find best libraries for task
/sota <topic>	Get state-of-the-art approaches

ML-Specific Commands

Command	Description
/experiment <name>	Start a new experiment
/metrics	Show current training metrics
/tensorboard	Open TensorBoard
/checkpoint	Save model state
/compare <e1> <e2>	Compare experiments
/huggingface <query>	Search HuggingFace Hub

Visual Mode

Launch with --visual to see all agents in a tmux layout:

┌───────────────────┬───────────────────┬───────────────────┐
│   ORCHESTRATOR    │     ENGINEER      │      TESTER       │
│                   │                   │                   │
│ Breaking down     │ Implementing      │ Waiting for       │
│ task into         │ auth module...    │ implementation... │
│ subtasks...       │                   │                   │
├───────────────────┼───────────────────┼───────────────────┤
│     SECURITY      │      SCOUT        │      [MAIN]       │
│                   │                   │                   │
│ Queued for        │ Searching JWT     │ You: _            │
│ review            │ best practices... │                   │
│                   │                   │                   │
└───────────────────┴───────────────────┴───────────────────┘

How It Works

                          ┌─────────────┐
                          │    User     │
                          └──────┬──────┘
                                 │
                                 ▼
                    ┌────────────────────────┐
                    │     Orchestrator       │
                    │  (task decomposition)  │
                    └───────────┬────────────┘
                                │
           ┌────────────────────┼────────────────────┐
           │                    │                    │
           ▼                    ▼                    ▼
   ┌───────────────┐   ┌───────────────┐   ┌───────────────┐
   │   Engineer    │   │    Tester     │   │   Security    │
   │ (implements)  │   │  (validates)  │   │  (reviews)    │
   └───────┬───────┘   └───────┬───────┘   └───────┬───────┘
           │                    │                    │
           └────────────────────┴────────────────────┘
                                │
                                ▼
                    ┌────────────────────────┐
                    │    File-Based IPC      │
                    │ workspace/tasks/*.json │
                    └────────────────────────┘

1. You enter a request in the interactive session
2. Orchestrator analyzes and breaks it into subtasks
3. Specialist agents spawn as parallel Claude subprocesses
4. Agents work on your actual project files
5. Results aggregate back through the orchestrator
6. You see the combined output with full context

Configuration

Create ~/.agentk/config.sh:

# Custom model (default: claude-3-sonnet)
export AGENTK_MODEL="claude-3-opus-20240229"

# Log level: debug, info, warn, error
export LOG_LEVEL="info"

# Custom workspace location
export AGENTK_WORKSPACE="/custom/path/workspace"

# Parallel agent limit (default: 4)
export AGENTK_MAX_PARALLEL=6

Project Structure

agentk/
├── agentk                # Main CLI entry point
├── lib/
│   ├── core.sh          # Core utilities, logging, date context
│   ├── ui.sh            # Terminal UI, spinners, chat boundaries
│   ├── ipc.sh           # Inter-process communication
│   ├── spawn.sh         # Agent subprocess management
│   └── visual.sh        # tmux integration
├── modes/
│   ├── shared/
│   │   └── scout.md     # Scout agent system prompt
│   ├── dev/             # Development mode prompts
│   │   ├── orchestrator.md
│   │   ├── engineer.md
│   │   ├── tester.md
│   │   └── security.md
│   └── ml/              # ML mode prompts
│       ├── orchestrator.md
│       ├── researcher.md
│       ├── ml-engineer.md
│       ├── data-engineer.md
│       └── evaluator.md
└── workspace/           # Runtime data (gitignored)
    ├── tasks/           # Task queue (JSON)
    ├── results/         # Agent outputs
    ├── logs/            # Agent logs
    └── experiments/     # ML experiment tracking

Known Limitations

Limitation	Workaround
File conflicts when agents edit same file	Use /focus to serialize work on critical files
Each agent = separate API call (cost)	Use orchestrator's judgment on when to parallelize
Agents don't share real-time context	Orchestrator maintains shared state in workspace
Rate limiting with many parallel agents	AGENTK_MAX_PARALLEL limits concurrent spawns

Roadmap

• Web UI dashboard
• Custom agent definitions
• Persistent conversation history
• Cost tracking per agent
• Team collaboration mode
• Plugin system for custom tools

Contributing

Contributions welcome!

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing)
3. Make your changes
4. Run tests (make test)
5. Submit a pull request

License

MIT License - see LICENSE for details.

Acknowledgments

• Inspired by Boris Cherny's parallel Claude workflow
• Built for the Claude Code community
• Powered by Anthropic's Claude

---

AGENT-K - Because one Claude is good, but a team of Claudes is better.

GitHub | npm | PyPI