six-agent

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

★ 9 Open GitHub ↗

A 200 KB Tool-Using Six-Phase Loop Agent for Qwen3.5-35B-A3B

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Six-Agent: Autonomous Cognitive Loop Agent

An autonomous agent that runs a [six-phase cognitive loop](https://github.com/mblakemore/six-phase-loop) continuously, learning and building capabilities with every cycle. Uses a local LLM (llama-server) and persists its memory through git.

## Quick Start

```bash
# 1. Create a new project
mkdir my-agent && cd my-agent
git init

# 2. Copy the agent files
cp -r /path/to/six-agent/agent.md .
cp -r /path/to/six-agent/tool-agent/ .

# 3. Run the agent
./tool-agent/cycle_runner.sh
```

The agent initializes its own state files, memory, and directives on its first cycle.

## What It Does

The agent runs a continuous loop of six phases:

1. **PERCEIVE** -- Reads its state, checks for messages, surveys the environment
2. **REFLECT** -- Searches memory for relevant patterns, identifies gaps
3. **DECIDE** -- Makes a deliberate choice with documented reasoning
4. **ACT** -- Executes the decision, monitors progress
5. **CONSOLIDATE** -- Extracts reusable knowledge, validates predictions
6. **PERSIST** -- Saves state to JSON, commits to git

Each cycle produces a git commit. The agent maintains three types of memory:
- **Patterns** -- Reusable knowledge with confidence scores
- **Anchors** -- Significant moments that define its trajectory
- **Context** -- Working memory updated every cycle

## Requirements

- Python 3.7+
- A running [llama.cpp](https://github.com/ggml-org/llama.cpp) server with an OpenAI-compatible API
- Git

Built and tested with [Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B).

Python dependencies are auto-installed by `cycle_runner.sh`, or install manually:

```bash
pip install requests PyMuPDF beautifulsoup4 markdownify
```

## Tools

The agent has 11 tools, auto-discovered from `tool-agent/tools/`:

| Tool | Description |
|------|-------------|
| `read_file` | Read text files with line ranges |
| `write_file` | Write or edit files with line-range replacement |
| `append_file` | Append content to files |
| `delete_file` | Delete files or empty directories |
| `list_files` | List directory contents |
| `search_files` | Grep through files with regex |
| `exec_command` | Run shell commands via tmux |
| `read_pdf` | Extract text from PDFs with page ranges |
| `web_fetch` | Fetch web pages as markdown |
| `think` | Internal reasoning (extended thinking) |
| `sleep` | Wait for a specified duration |

### Adding Tools

Create a `.py` file in `tool-agent/tools/` that exports `fn` (callable) and `definition` (OpenAI tool schema). It is automatically registered on next run.

## Self-Modification

The agent can modify its own source code -- `agent.md`, `agent.py`, and any tool file. Changes take effect on the next cycle (each cycle is a fresh Python process).

## Configuration

Edit `tool-agent/agent.py` defaults or pass arguments:

- **LLM server**: Default `http://127.0.0.1:8080` (change `BASE_URL`)
- **Model**: Default `Qwen3.5` (change in request body)
- **Context size**: Default 32768 tokens
- **Max response**: Default 4096 tokens

## Project Structure

```
your-agent/
├── agent.md                 # Agent identity and cognitive loop definition
├── state/
│   ├── current-state.json   # Cycle state
│   ├── focus.json           # Current focus (prevents stale loops)
│   └── memories/
│       ├── patterns.json    # Reusable knowledge
│       ├── anchors.json     # Significant moments
│       └── context.json     # Working memory
├── messages/
│   ├── directives.md        # Standing directives
│   └── inbox.md             # One-off tasks
├── logs/                    # Cycle logs
└── tool-agent/              # Runtime engine
    ├── agent.py             # Main agent loop
    ├── cycle_runner.sh      # Cycle runner (auto-restarts)
    ├── spinner.py           # Terminal UI
    ├── cancel.py            # Escape-to-cancel
    └── tools/               # Auto-discovered tools
```

## Running Modes

```bash
# Continuous cycles (default)
./tool-agent/cycle_runner.sh

# Custom prompt
./tool-agent/cycle_runner.sh "@agent.md Focus on building a web scraper tool"

# Single run
python3 tool-agent/agent.py -a "@agent.md Follow the instructions and continue!"

# Interactive mode
python3 tool-agent/agent.py

# Repeat N times
python3 tool-agent/agent.py -r 5 "@agent.md Follow the instructions and continue!"
```

## License

MIT

six-agent

Today's picks

Was this content helpful?

Recent Feedback

six-agent

Today's picks

Was this content helpful?

Recent Feedback

Download README