**Version:** 0.1.14
# When Simplicity Becomes Your Superpower: Meet KISS Multi Agent Evolutionary Framework
*"Everything should be made as simple as possible, but not simpler." — Albert Einstein*
______________________________________________________________________
KISS stands for ["Keep it Simple, Stupid"](https://en.wikipedia.org/wiki/KISS_principle) which is a well known software engineering principle.
## 🎯 The Problem with AI Agent Frameworks Today
Let's be honest. The AI agent ecosystem has become a jungle.
Every week brings a new framework promising to revolutionize how we build AI agents. They come loaded with abstractions on top of abstractions, configuration files that rival tax forms, and dependency trees that make `node_modules` look tidy. By the time you've figured out how to make your first tool call, you've already burned through half your patience and all your enthusiasm.
**What if there was another way?**
What if building AI agents could be as straightforward as the name suggests?
Enter **KISS** — the *Keep It Simple, Stupid* Agent Framework.
## 🚀 Your First Agent in 30 Seconds.
Try the interactive Jupyter notebook by running `uv run notebook --lab`.
Let me show you something beautiful:
```python
from kiss.core.kiss_agent import KISSAgent
def calculate(expression: str) -> str:
"""Evaluate a math expression."""
return str(eval(expression))
agent = KISSAgent(name="Math Buddy")
result = agent.run(
model_name="gemini-2.5-flash",
prompt_template="Calculate: {question}",
arguments={"question": "What is 15% of 847?"},
tools=[calculate]
)
print(result) # 127.05
```
That's a fully functional AI agent that uses tools. No annotations. No boilerplate. No ceremony. Just intent, directly expressed.
KISS uses **native function calling** from the LLM providers for efficiency and accuracy. Your Python functions become tools automatically. Type hints become schemas. Docstrings become descriptions. No crazy annotations. Everything just works. Well you might ask "**why not use LangChain, DSpy, OpenHands, MiniSweAgent, CrewAI, Google ADK, Claude Agent SDK, or some well established agent frameworks?**" Here is my response:
- **KISS comes with [Repo Optimizer](src/kiss/agents/coding_agents/repo_optimizer.py) that will not only enable you write or create agents, but also automatically optimize the agents for efficiency and cost.**
- **It has the GEPA prompt optimizer builtin with a simple API.**
- **It has a [RelentlessCodingAgent](src/kiss/agents/coding_agents/relentless_coding_agent.py), which is pretty straightforward in terms of implementation, but it can work for very very long tasks. It was self evolved over time and is still evolving.**
- **The framework can self optimize based on your requirements (e.g. maximize test coverage).**
- **No bloat and simple codebase.**
- **Optimization strategies can be written as text.**
## 🤝 Multi-Agent Orchestration is Function Composition
Here's where KISS really shines — composing multiple agents into systems greater than the sum of their parts.
Since agents are just functions, you orchestrate them with plain Python. Here's a complete **research-to-article pipeline** with three agents:
```python
from kiss.core.kiss_agent import KISSAgent
# Agent 1: Research a topic
researcher = KISSAgent(name="Researcher")
research = researcher.run(
model_name="gpt-4o",
prompt_template="List 3 key facts about {topic}. Be concise.",
arguments={"topic": "Python asyncio"},
is_agentic=False # Simple generation, no tools
)
# Agent 2: Write a draft using the research
writer = KISSAgent(name="Writer")
draft = writer.run(
model_name="claude-sonnet-4-5",
prompt_template="Write a 2-paragraph intro based on:\n{research}",
arguments={"research": research},
is_agentic=False
)
# Agent 3: Polish the draft
editor = KISSAgent(name="Editor")
final = editor.run(
model_name="gemini-2.5-flash",
prompt_template="Improve clarity and fix any errors:\n{draft}",
arguments={"draft": draft},
is_agentic=False
)
print(final)
```
**That's it.** Each agent can use a different model. Each agent saves its own trajectory. And you compose them with the most powerful orchestration tool ever invented: **regular Python code**.
No special orchestration framework needed. No message buses. No complex state machines. Just Python functions calling Python functions.
## 💪 Using Relentless Coding Agent
For very long running coding tasks, use the `RelentlessCodingAgent`. The agent will work relentlessly to complete your task using a single-agent architecture with smart continuation:
```python
from kiss.agents.coding_agents.relentless_coding_agent import RelentlessCodingAgent
agent = RelentlessCodingAgent(name="Simple Coding Agent")
result = agent.run(
prompt_template="""
Create a Python script that reads a CSV file,
filters rows where age > 18, and writes to a new file.
""",
model_name="claude-sonnet-4-5",
work_dir="./workspace",
max_steps=200,
max_sub_sessions=200
)
print(f"Result: {result}")
```
**Running with Docker:**
You can optionally run bash commands inside a Docker container for isolation:
```python
from kiss.agents.coding_agents.relentless_coding_agent import RelentlessCodingAgent
agent = RelentlessCodingAgent(name="Dockered Relentless Coding Agent")
result = agent.run(
prompt_template="""
Install numpy and create a script that generates
a random matrix and computes its determinant.
""",
docker_image="python:3.11-slim", # Bash commands run in Docker
max_steps=200,
max_sub_sessions=2000
)
print(f"Result: {result}")
```
**Key Features:**
- **Single-Agent with Auto-Continuation**: A single agent executes the task across multiple sub-sessions, automatically continuing where it left off via structured JSON progress tracking
- **Structured Progress Tracking**: Each sub-session reports completed and remaining tasks in JSON format (done/next items), which is deduplicated and passed to subsequent sub-sessions along with a scan of existing files in the work directory
- **Compressed Prompts**: Minimal, high-signal task prompts with critical rules (use Write() for new files, bounded poll loops, immediate finish on success)
- **Efficiency Rules**: Built-in prompt instructions enforce immediate completion when tests pass, timeout guidance for bash, and bounded loops for background jobs
- **Retry with Context**: Failed sub-sessions automatically pass structured progress summaries and file listings to the next sub-session
- **Configurable Sub-Sessions**: Set high sub-session counts (e.g., 200+) for truly relentless execution
- **Docker Support**: Optional isolated execution via Docker containers
- **Path Access Control**: Enforces read/write permissions on file system paths
- **Built-in Tools**: Bash, Read, Edit, and Write tools for file operations
- **Budget & Token Tracking**: Automatic cost and token usage monitoring across all sub-sessions
## 🔧 Using Repo Optimizer
The `RepoOptimizer` (`repo_optimizer.py`) uses the `RelentlessCodingAgent` to optimize code within your own project repository. It runs the target program, monitors output in real time, fixes errors, and iteratively optimizes for speed and cost — all without changing the agent's interface.
```python
from kiss.agents.coding_agents.repo_optimizer import main
# Run the repo optimizer (uses RelentlessCodingAgent with claude-opus-4-6)
main()
```
Or run it directly from the command line:
```bash
uv run python -m kiss.agents.coding_agents.repo_optimizer
```
**How It Works:**
1. Runs the target program (e.g., `relentless_coding_agent.py`) and monitors output in real time
1. If repeated errors are observed, fixes them and reruns
1. Once the program succeeds, analyzes output and optimizes the source for speed and cost
1. Repeats until running time and cost are reduced significantly
**Optimization Strategies:**
- Shorter system prompts preserving meaning
- Removing redundant instructions and minimizing conversation turns
- Batching operations and using early termination
- Applying latest agentic patterns for long-horizon tasks
- Inventing and implementing new agent architectures for efficiency and reliability
## 🎨 Output Formatting
Unlike other agentic systems, you do not need to specify the output schema for the agent. Just create
a suitable "finish" function with parameters. The parameters could be treated as the top level keys
in a json format.
**Example: Custom Structured Output**
```python
from kiss.core.kiss_agent import KISSAgent
# Define a custom finish function with your desired output structure
def finish(
sentiment: str,
confidence: float,
key_phrases: str,
summary: str
) -> str:
"""
Complete the analysis with structured results.
Args:
sentiment: The overall sentiment ('positive', 'negative', or 'neutral')
confidence: Confidence score between 0.0 and 1.0
key_phrases: Comma-separated list of key phrases found in the text
summary: A brief summary of the analysis
Returns:
The formatted analysis result
"""
...
```
The agent will automatically use your custom `finish` function instead of the default one which returns its argument. The function's parameters define what information the agent must provide, and the docstring helps the LLM understand how to format each field.
## 📊 Trajectory Saving and Visualization
Agent trajectories are automatically saved to the artifacts directory (default: `artifacts/`). Each trajectory includes:
- Complete message history with token usage and budget information appended to each message
- Tool calls and results
- Configuration used
- Timestamps
- Budget and token usage statistics
### Visualizing Trajectories
The framework includes a web-based trajectory visualizer for viewing agent execution histories:
```bash
# Run the visualizer server
uv run python -m kiss.viz_trajectory.server artifacts
# Or with custom host/port
uv run python -m kiss.viz_trajectory.server artifacts --host 127.0.0.1 --port 5050
```
Then open your browser to `http://127.0.0.1:5050` to view the trajectories.
The visualizer provides:
- **Modern UI**: Dark theme with smooth animations
- **Sidebar Navigation**: List of all trajectories sorted by start time
- **Markdown Rendering**: Full markdown support for message content
- **Code Highlighting**: Syntax highlighting for fenced code blocks
- **Message Display**: Clean, organized view of agent conversations
- **Metadata Display**: Shows agent ID, model, steps, tokens, and budget information
📖 **For detailed trajectory visualizer documentation, see [Trajectory Visualizer README](src/kiss/viz_trajectory/README.md)**
## 📖 Overview
KISS is a lightweight, yet powerful, multi agent framework that implements a ReAct (Reasoning and Acting) loop for LLM agents. The framework provides:
- **Simple Architecture**: Clean, minimal core that's easy to understand and extend
- **Relentless Coding Agent**: Single-agent coding system with smart auto-continuation for long-running tasks
- **Repo Optimizer**: Uses RelentlessCodingAgent to iteratively optimize code in your project for speed and cost (💡 new idea)
- **GEPA Implementation From Scratch**: Genetic-Pareto prompt optimization for compound AI systems
- **KISSEvolve Implementation From Scratch**: Evolutionary algorithm discovery framework with LLM-guided mutation and crossover
- **Model Agnostic**: Support for multiple LLM providers (OpenAI, Anthropic, Gemini, Together AI, OpenRouter)
- **Native Function Calling**: Seamless tool integration using native function calling APIs (OpenAI, Anthropic, Gemini, Together AI, and OpenRouter)
- **Docker Integration**: Built-in Docker manager for running agents in isolated environments
- **Trajectory Tracking**: Automatic saving of agent execution trajectories with unified state management
- **Token Streaming**: Real-time token streaming via async callback for all providers (OpenAI, Anthropic, Gemini, Together AI, OpenRouter), including thinking/reasoning tokens and tool execution output
- **Token Usage Tracking**: Built-in token usage tracking with automatic context length detection and step counting
- **Budget Tracking**: Automatic cost tracking and budget monitoring across all agent runs
- **Self-Evolution**: Framework for agents to evolve and refine other multi agents
- **SWE-bench Dataset Support**: Built-in support for downloading and working with SWE-bench Verified dataset
- **RAG Support**: Simple retrieval-augmented generation system with in-memory vector store
- **Useful Agents**: Pre-built utility agents including prompt refinement and general bash execution agents
- **Multiprocessing Support**: Utilities for parallel execution of functions using multiprocessing
- **Trajectory Visualization**: Web-based visualizer for viewing agent execution trajectories with modern UI
## 📦 Installation
```bash
# Install uv if you haven't already
curl -LsSf https://astral.sh/uv/install.sh | sh
# Clone/download KISS and navigate to the directory
cd kiss
# Create virtual environment
uv venv --python 3.13
# Install all dependencies (full installation)
uv sync
# (Optional) activate the venv for convenience (uv run works without activation)
source .venv/bin/activate
# Set up API keys (optional, for LLM providers)
export GEMINI_API_KEY="your-key-here"
export OPENAI_API_KEY="your-key-here"
export ANTHROPIC_API_KEY="your-key-here"
export TOGETHER_API_KEY="your-key-here"
export OPENROUTER_API_KEY="your-key-here"
```
### Selective Installation (Dependency Groups)
KISS supports selective installation via dependency groups for minimal footprints:
```bash
# Minimal core only (no model SDKs) - for custom integrations
uv sync --group core
# Core + specific provider support
uv sync --group claude # Core + Anthropic Claude
uv sync --group openai # Core + OpenAI Compatible Models
uv sync --group gemini # Core + Google Gemini
# Docker support (for running agents in isolated containers)
uv sync --group docker
# Evals dependencies (for running benchmarks)
uv sync --group evals
# Development tools (mypy, ruff, pytest, jupyter, etc.)
uv sync --group dev
# Combine multiple groups as needed
uv sync --group claude --group dev
```
**Dependency Group Contents:**
| Group | Description | Key Packages |
|-------|-------------|--------------|
| `core` | Minimal core module | pydantic, rich, requests, beautifulsoup4, playwright, flask |
| `claude` | Core + Anthropic | core + anthropic |
| `openai` | Core + OpenAI | core + openai |
| `gemini` | Core + Google | core + google-genai |
| `docker` | Docker integration | docker, types-docker |
| `evals` | Benchmark running | datasets, swebench, orjson, scipy, scikit-learn |
| `dev` | Development tools | mypy, ruff, pyright, pytest, jupyter, notebook |
> **Optional Dependencies:** All LLM provider SDKs (`openai`, `anthropic`, `google-genai`) are optional. You can import `kiss.core` and `kiss.agents` without installing all of them. When you try to use a model whose SDK is not installed, KISS raises a clear `KISSError` telling you which package to install.
## 📚 KISSAgent API Reference
📖 **For detailed KISSAgent API documentation, see [API.md](API.md)**
## 🎯 Using GEPA for Prompt Optimization
KISS has a fresh implementation of GEPA with some improvements. GEPA (Genetic-Pareto) is a prompt optimization framework that uses natural language reflection to evolve prompts. It maintains an instance-level Pareto frontier of top-performing prompts and combines complementary lessons through structural merge. GEPA is based on the paper ["GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning"](https://arxiv.org/pdf/2507.19457).
📖 **For detailed GEPA documentation, see [GEPA README](src/kiss/agents/gepa/README.md)**
## 🧪 Using KISSEvolve for Algorithm Discovery
KISSEvolve is an evolutionary algorithm discovery framework that uses LLM-guided mutation and crossover to evolve code variants. It supports advanced features including island-based evolution, novelty rejection sampling, and multiple parent sampling methods.
For usage examples, API reference, and configuration options, please see the [KISSEvolve README](src/kiss/agents/kiss_evolve/README.md).
📖 **For detailed KISSEvolve documentation, see [KISSEvolve README](src/kiss/agents/kiss_evolve/README.md)**
## 🔄 Using Self-Evolving Multi-Agent
The Self-Evolving Multi-Agent is an advanced coding agent with planning, error recovery, dynamic tool creation, and the ability to evolve itself for better efficiency and accuracy using KISSEvolve.
📖 **For detailed Self-Evolving Multi-Agent documentation, see [Self-Evolving Multi-Agent README](src/kiss/agents/self_evolving_multi_agent/README.md)**
```python
from kiss.agents.self_evolving_multi_agent import SelfEvolvingMultiAgent
# Create and run the agent
agent = SelfEvolvingMultiAgent()
result = agent.run("""
Create a Python script that:
1. Generates the first 20 Fibonacci numbers
2. Saves them to a file called 'fibonacci.txt'
3. Reads the file back and prints the sum
""")
print(result)
# Access execution statistics
stats = agent.get_stats()
print(f"Completed todos: {stats['completed']}/{stats['total_todos']}")
print(f"Dynamic tools created: {stats['dynamic_tools']}")
```
**Key Features:**
- **Planning & Task Tracking**: Creates and manages a todo list with status tracking (pending → in_progress → completed/failed)
- **Sub-Agent Delegation**: Spawns focused sub-agents for individual task execution
- **Dynamic Tool Creation**: Creates reusable tools at runtime when prompted by the orchestrator
- **Error Recovery**: Automatic retry logic with configurable max retries
- **Docker Isolation**: Runs code execution in isolated Docker containers
- **Self-Evolution**: Uses KISSEvolve to optimize for efficiency and accuracy
**Evolving the Agent:**
```python
from kiss.agents.self_evolving_multi_agent.agent_evolver import AgentEvolver
# Create evolver
evolver = AgentEvolver(
package_name="kiss.agents.self_evolving_multi_agent",
agent_file_path="multi_agent.py",
model_name="gemini-3-flash-preview",
focus_on_efficiency=True,
)
# Run baseline evaluation first
baseline = evolver.run_baseline_evaluation()
print(f"Baseline fitness: {baseline['fitness']:.4f}")
# Evolve the agent
best = evolver.evolve()
print(f"Evolved fitness: {best.fitness:.4f}")
# Save the best variant
evolver.save_best(best)
```
## 🔍 Using SimpleRAG for Retrieval-Augmented Generation
SimpleRAG provides a lightweight RAG system with in-memory vector storage and similarity search:
> **Note**: SimpleRAG requires a model with embedding support. Currently, OpenAI, Together AI, and Gemini models support embeddings. Anthropic models do not provide embedding APIs.
```python
from kiss.rag import SimpleRAG
# Initialize RAG system with a model name that supports embeddings
rag = SimpleRAG(model_name="gpt-4o", metric="cosine") # or "l2" for L2 distance
# Add documents
documents = [
{
"id": "1",
"text": "Python is a programming language known for its simplicity.",
"metadata": {"topic": "programming", "language": "Python"},
},
{
"id": "2",
"text": "Machine learning uses algorithms to learn from data.",
"metadata": {"topic": "ML", "field": "AI"},
},
{
"id": "3",
"text": "Docker containers provide isolated execution environments.",
"metadata": {"topic": "devops", "tool": "Docker"},
},
]
rag.add_documents(documents)
# Query similar documents
results = rag.query("What is Python?", top_k=2)
for result in results:
print(f"ID: {result['id']}")
print(f"Text: {result['text']}")
print(f"Score: {result['score']:.4f}")
print(f"Metadata: {result['metadata']}")
print()
# Query with filter
def filter_by_topic(doc: dict) -> bool:
return doc.get("metadata", {}).get("topic") == "programming"
filtered_results = rag.query("programming language", top_k=5, filter_fn=filter_by_topic)
# Get collection statistics
stats = rag.get_collection_stats()
print(f"Documents: {stats['num_documents']}, Embedding dim: {stats['embedding_dimension']}")
# Delete documents
rag.delete_documents(["1", "2"])
# Get a specific document
doc = rag.get_document("3")
# Clear all documents
rag.clear_collection()
```
## ⚡ Multiprocessing
KISS provides utilities for parallel execution of Python functions using multiprocessing. This is useful for running multiple independent tasks concurrently to maximize CPU utilization.
### Basic Usage
```python
from kiss.multiprocessing import run_functions_in_parallel
def add(a, b):
return a + b
def multiply(x, y):
return x * y
# Define tasks as (function, arguments) tuples
tasks = [(add, [1, 2]), (multiply, [3, 4])]
results = run_functions_in_parallel(tasks)
print(results) # [3, 12]
```
### With Keyword Arguments
```python
from kiss.multiprocessing import run_functions_in_parallel_with_kwargs
def greet(name, title="Mr."):
return f"Hello, {title} {name}!"
functions = [greet, greet]
args_list = [["Alice"], ["Bob"]]
kwargs_list = [{"title": "Dr."}, {}]
results = run_functions_in_parallel_with_kwargs(functions, args_list, kwargs_list)
print(results) # ["Hello, Dr. Alice!", "Hello, Mr. Bob!"]
```
### 💻 Checking Available Cores
```python
from kiss.multiprocessing import get_available_cores
num_cores = get_available_cores()
print(f"Available CPU cores: {num_cores}")
```
The multiprocessing utilities automatically scale to the number of available CPU cores, using at most as many workers as there are tasks to avoid unnecessary overhead.
## 🐳 Docker Manager
KISS provides a `DockerManager` class for managing Docker containers and executing commands inside them. This is useful for running code in isolated environments, testing with specific dependencies, or working with SWE-bench tasks.
### Basic Usage
```python
from kiss.docker import DockerManager
# Create a Docker manager for an Ubuntu container
with DockerManager(image_name="ubuntu", tag="22.04", workdir="/app") as docker:
# Run commands inside the container
output = docker.run_bash_command("echo 'Hello from Docker!'", "Print greeting")
print(output)
output = docker.run_bash_command("python3 --version", "Check Python version")
print(output)
```
### Manual Lifecycle Management
```python
from kiss.docker import DockerManager
docker = DockerManager(image_name="python", tag="3.11", workdir="/workspace")
docker.open() # Pull image and start container
try:
output = docker.run_bash_command("pip install numpy", "Install numpy")
output = docker.run_bash_command("python -c 'import numpy; print(numpy.__version__)'", "Check numpy")
print(output)
finally:
docker.close() # Stop and remove container
```
### Port Mapping
```python
from kiss.docker import DockerManager
# Map container port 8080 to host port 8080
with DockerManager(image_name="nginx", ports={80: 8080}) as docker:
# Start a web server
docker.run_bash_command("nginx", "Start nginx")
# Get the actual host port (useful when Docker assigns a random port)
host_port = docker.get_host_port(80)
print(f"Server available at http://localhost:{host_port}")
```
### Configuration Options
- `image_name`: Docker image name (e.g., 'ubuntu', 'python:3.11')
- `tag`: Image tag/version (default: 'latest')
- `workdir`: Working directory inside the container (default: '/')
- `mount_shared_volume`: Whether to mount a shared volume for file transfer (default: True)
- `ports`: Port mapping from container to host (e.g., `{8080: 8080}`)
The Docker manager automatically handles image pulling, container lifecycle, and cleanup of temporary directories.
## 📁 Project Structure
```
kiss/
├── src/kiss/
│ ├── agents/ # Agent implementations
│ │ ├── create_and_optimize_agent/ # Agent evolution and improvement
│ │ │ ├── agent_evolver.py # Evolutionary agent optimization
│ │ │ ├── improver_agent.py # Agent improvement through generations
│ │ │ ├── config.py # Agent creator configuration
│ │ │ ├── BLOG.md # Blog post about agent evolution
│ │ │ └── README.md # Agent creator documentation
│ │ ├── gepa/ # GEPA (Genetic-Pareto) prompt optimizer
│ │ │ ├── gepa.py
│ │ │ ├── config.py # GEPA configuration
│ │ │ └── README.md # GEPA documentation
│ │ ├── kiss_evolve/ # KISSEvolve evolutionary algorithm discovery
│ │ │ ├── kiss_evolve.py
│ │ │ ├── novelty_prompts.py # Prompts for novelty-based evolution
│ │ │ ├── config.py # KISSEvolve configuration
│ │ │ └── README.md # KISSEvolve documentation
│ │ ├── coding_agents/ # Coding agents for software development tasks
│ │ │ ├── relentless_coding_agent.py # Single-agent system with smart auto-continuation
│ │ │ ├── claude_coding_agent.py # Claude-based coding agent
│ │ │ ├── repo_optimizer.py # Iterative code optimizer using RelentlessCodingAgent
│ │ │ ├── repo_agent.py # Repo-level task agent using RelentlessCodingAgent
│ │ │ └── config.py # Coding agent configuration (RelentlessCodingAgent)
│ │ ├── self_evolving_multi_agent/ # Self-evolving multi-agent system
│ │ │ ├── agent_evolver.py # Agent evolution logic
│ │ │ ├── multi_agent.py # Multi-agent orchestration
│ │ │ ├── config.py # Configuration
│ │ │ └── README.md # Documentation
│ │ └── kiss.py # Utility agents (prompt refiner, bash agent)
│ ├── core/ # Core framework components
│ │ ├── base.py # Base class with common functionality for all KISS agents
│ │ ├── kiss_agent.py # KISS agent with native function calling
│ │ ├── printer.py # Abstract Printer base class and MultiPrinter
│ │ ├── print_to_console.py # ConsolePrinter: Rich-formatted terminal output
│ │ ├── print_to_browser.py # BrowserPrinter: SSE streaming to browser UI
│ │ ├── config.py # Configuration
│ │ ├── config_builder.py # Dynamic config builder with CLI support
│ │ ├── kiss_error.py # Custom error class
│ │ ├── utils.py # Utility functions (finish, resolve_path, is_subpath, etc.)
│ │ ├── useful_tools.py # UsefulTools class with path-restricted Read, Write, Bash, Edit, search_web, fetch_url
│ │ └── models/ # Model implementations
│ │ ├── model.py # Model interface with TokenCallback streaming support
│ │ ├── gemini_model.py # Gemini model implementation
│ │ ├── openai_compatible_model.py # OpenAI-compatible API model (OpenAI, Together AI, OpenRouter)
│ │ ├── anthropic_model.py # Anthropic model implementation
│ │ └── model_info.py # Model info: context lengths, pricing, and capabilities
│ ├── docker/ # Docker integration
│ │ └── docker_manager.py
│ ├── evals/ # Benchmark and evaluation integrations
│ │ ├── algotune/ # AlgoTune benchmark integration
│ │ │ ├── run_algotune.py # AlgoTune task evolution
│ │ │ └── config.py # AlgoTune configuration
│ │ ├── arvo_agent/ # ARVO vulnerability detection agent
│ │ │ ├── arvo_agent.py # Arvo-based vulnerability detector
│ │ │ └── arvo_tags.json # Docker image tags for Arvo
│ │ ├── hotpotqa/ # HotPotQA benchmark integration
│ │ │ ├── hotpotqa_benchmark.py # HotPotQA benchmark runner
│ │ │ └── README.md # HotPotQA documentation
│ │ └── swe_agent_verified/ # SWE-bench Verified benchmark integration
│ │ ├── run_swebench.py # Main runner with CLI support
│ │ ├── config.py # Configuration for SWE-bench runs
│ │ └── README.md # SWE-bench documentation
│ ├── multiprocessing/ # Multiprocessing utilities
│ │ └── multiprocess.py
│ ├── rag/ # RAG (Retrieval-Augmented Generation)
│ │ └── simple_rag.py # Simple RAG system with in-memory vector store
│ ├── demo/ # Demo scripts
│ │ └── kiss_demo.py # Interactive demo with streaming output to terminal and browser
│ ├── scripts/ # Utility scripts
│ │ ├── check.py # Code quality check script
│ │ ├── notebook.py # Jupyter notebook launcher and utilities
│ │ ├── kissevolve_bubblesort.py # KISSEvolve example: evolving bubble sort
│ │ └── release.sh # Release automation script
│ ├── tests/ # Test suite
│ │ ├── conftest.py # Pytest configuration and fixtures
│ │ ├── test_kiss_agent_agentic.py
│ │ ├── test_kiss_agent_non_agentic.py
│ │ ├── test_kissevolve_bubblesort.py
│ │ ├── test_gepa_hotpotqa.py
│ │ ├── test_gepa_progress_callback.py # Tests for GEPA progress callbacks
│ │ ├── test_docker_manager.py
│ │ ├── test_model_implementations.py # Integration tests for model implementations
│ │ ├── run_all_models_test.py # Comprehensive tests for all models
│ │ ├── test_multiprocess.py
│ │ ├── test_internal.py
│ │ ├── test_core_branch_coverage.py # Branch coverage tests for core components
│ │ ├── test_gemini_model_internals.py # Tests for Gemini model internals
│ │ ├── test_cli_options.py # Tests for CLI option parsing
│ │ ├── test_claude_coding_agent.py # Tests for coding agents
│ │ ├── test_evolver_progress_callback.py # Tests for AgentEvolver progress callbacks
│ │ ├── test_token_callback.py # Tests for async token streaming callback
│ │ ├── test_coding_agent_token_callback.py # Tests for token callback in coding agents
│ │ ├── test_a_model.py # Tests for model implementations
│ │ ├── test_print_to_console.py # Tests for ConsolePrinter output
│ │ ├── test_print_to_browser.py # Tests for BrowserPrinter browser output
│ │ ├── test_search_web.py
│ │ └── test_useful_tools.py
│ ├── py.typed # PEP 561 marker for type checking
│ └── viz_trajectory/ # Trajectory visualization
│ ├── server.py # Flask server for trajectory visualization
│ ├── README.md # Trajectory visualizer documentation
│ └── templates/ # HTML templates for the visualizer
│ └── index.html
├── scripts/ # Repository-level scripts
│ └── release.sh # Release script
├── pyproject.toml # Project configuration
└── README.md
```
## 🏷️ Versioning
The project uses semantic versioning (MAJOR.MINOR.PATCH). The version is defined in a single source of truth:
- **Version file**: `src/kiss/_version.py` - Edit this file to update the version
- **Package access**: `kiss.__version__` - Access the version programmatically
- **Build system**: `pyproject.toml` automatically reads the version from `_version.py` using dynamic versioning
Example:
```python
from kiss import __version__
print(f"KISS version: {__version__}")
```
To update the version, simply edit `src/kiss/_version.py`:
```python
__version__ = "0.2.0" # Update to new version
```
## ⚙️ Configuration
Configuration is managed through environment variables and the `DEFAULT_CONFIG` object:
- **API Keys**: Set `GEMINI_API_KEY`, `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `TOGETHER_API_KEY`, and/or `OPENROUTER_API_KEY` environment variables
- **Agent Settings**: Modify `DEFAULT_CONFIG.agent` in `src/kiss/core/config.py`:
- `max_steps`: Maximum iterations in the ReAct loop (default: 100)
- `verbose`: Enable verbose output (default: True)
- `debug`: Enable debug mode (default: False)
- `max_agent_budget`: Maximum budget per agent run in USD (default: 10.0)
- `global_max_budget`: Maximum total budget across all agents in USD (default: 200.0)
- `use_web`: Automatically add web browsing and search tool if enabled (default: True)
- `print_to_console`: Enable ConsolePrinter for Rich terminal output (default: True). Can be overridden per-call via the `print_to_console` parameter on `run()`.
- `print_to_browser`: Enable BrowserPrinter for live browser UI output (default: False). Can be overridden per-call via the `print_to_browser` parameter on `run()`.
- `artifact_dir`: Directory for agent artifacts (default: auto-generated with timestamp)
- **Relentless Coding Agent Settings**: Modify `DEFAULT_CONFIG.coding_agent.relentless_coding_agent` in `src/kiss/agents/coding_agents/config.py`:
- `model_name`: Model for task execution (default: "claude-opus-4-6")
- `max_sub_sessions`: Maximum number of sub-sessions for auto-continuation (default: 200)
- `max_steps`: Maximum steps per sub-session (default: 200)
- `max_budget`: Maximum budget in USD (default: 200.0)
- **GEPA Settings**: Modify `DEFAULT_CONFIG.gepa` in `src/kiss/agents/gepa/config.py`:
- `reflection_model`: Model to use for reflection (default: "gemini-3-flash-preview")
- `max_generations`: Maximum number of evolutionary generations (default: 10)
- `population_size`: Number of candidates to maintain in population (default: 8)
- `pareto_size`: Maximum size of Pareto frontier (default: 4)
- `mutation_rate`: Probability of mutating a prompt template (default: 0.5)
- **KISSEvolve Settings**: Modify `DEFAULT_CONFIG.kiss_evolve` in `src/kiss/agents/kiss_evolve/config.py`:
- `max_generations`: Maximum number of evolutionary generations (default: 10)
- `population_size`: Number of variants to maintain in population (default: 8)
- `mutation_rate`: Probability of mutating a variant (default: 0.7)
- `elite_size`: Number of best variants to preserve each generation (default: 2)
- `num_islands`: Number of islands for island-based evolution, 1 = disabled (default: 2)
- `migration_frequency`: Number of generations between migrations (default: 5)
- `migration_size`: Number of individuals to migrate between islands (default: 1)
- `migration_topology`: Migration topology: 'ring', 'fully_connected', or 'random' (default: "ring")
- `enable_novelty_rejection`: Enable code novelty rejection sampling (default: False)
- `novelty_threshold`: Cosine similarity threshold for rejecting code (default: 0.95)
- `max_rejection_attempts`: Maximum rejection attempts before accepting (default: 5)
- `parent_sampling_method`: Parent sampling: 'tournament', 'power_law', or 'performance_novelty' (default: "power_law")
- `power_law_alpha`: Power-law sampling parameter for rank-based selection (default: 1.0)
- `performance_novelty_lambda`: Selection pressure parameter for sigmoid (default: 1.0)
- **Self-Evolving Multi-Agent Settings**: Modify `DEFAULT_CONFIG.self_evolving_multi_agent` in `src/kiss/agents/self_evolving_multi_agent/config.py`:
- `model`: LLM model to use for the main agent (default: "gemini-3-flash-preview")
- `sub_agent_model`: Model for sub-agents (default: "gemini-3-flash-preview")
- `evolver_model`: Model for evolution (default: "gemini-3-flash-preview")
- `max_steps`: Maximum orchestrator steps (default: 100)
- `max_budget`: Maximum budget in USD (default: 10.0)
- `max_retries`: Maximum retries on error (default: 3)
- `sub_agent_max_steps`: Maximum steps for sub-agents (default: 50)
- `sub_agent_max_budget`: Maximum budget for sub-agents in USD (default: 2.0)
- `docker_image`: Docker image for execution (default: "python:3.12-slim")
- `workdir`: Working directory in container (default: "/workspace")
## 🛠️ Available Commands
### Development
- `uv sync` - Install all dependencies (full installation)
- `uv sync --group dev` - Install dev tools (mypy, ruff, pytest, jupyter, etc.)
- `uv sync --group <name>` - Install specific dependency group (see [Selective Installation](#selective-installation-dependency-groups))
- `uv build` - Build the project package
### Testing
- `uv run pytest` - Run all tests (uses testpaths from pyproject.toml)
- `uv run pytest src/kiss/tests/ -v` - Run all tests with verbose output
- `uv run pytest src/kiss/tests/test_kiss_agent_agentic.py -v` - Run agentic agent tests
- `uv run pytest src/kiss/tests/test_kiss_agent_non_agentic.py -v` - Run non-agentic agent tests
- `uv run pytest src/kiss/tests/test_multiprocess.py -v` - Run multiprocessing tests
- `uv run python -m unittest src.kiss.tests.test_docker_manager -v` - Run docker manager tests (unittest)
- `uv run python -m unittest discover -s src/kiss/tests -v` - Run all tests using unittest
### Code Quality
- `uv run check` - Run all code quality checks (fresh dependency install, build, lint, and type check)
- `uv run check --clean` - Run all code quality checks (fresh dependency install, build, lint, and type check after removing previous build options)
- `uv run ruff format src/` - Format code with ruff (line-length: 100, target: py313)
- `uv run ruff check src/` - Lint code with ruff (selects: E, F, W, I, N, UP)
- `uv run mypy src/` - Type check with mypy (python_version: 3.13)
- `uv run pyright src/` - Type check with pyright (alternative to mypy, stricter checking)
### Notebook
- `uv run notebook --test` - Test all imports and basic functionality
- `uv run notebook --lab` - Open the tutorial notebook in JupyterLab (recommended)
- `uv run notebook --run` - Open the tutorial notebook in Jupyter Notebook
- `uv run notebook --execute` - Execute notebook cells and update outputs in place
- `uv run notebook --convert` - Convert notebook to Python script
### Cleanup
```bash
rm -rf build/ dist/ .pytest_cache .mypy_cache .ruff_cache && \
find . -type d -name __pycache__ -exec rm -r {} + && \
find . -type f -name "*.pyc" -delete
```
## 🤖 Models Supported
**Supported Models**: The framework includes context length, pricing, and capability flags for:
**Generation Models** (text generation with function calling support):
- **OpenAI**: gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-4o, gpt-4o-mini, gpt-4-turbo, gpt-4, gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, gpt-5.1, gpt-5.2, gpt-5.2-pro
- **OpenAI (Codex)**: gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, gpt-5.2-codex, codex-mini-latest
- **OpenAI (Reasoning)**: o1, o1-mini, o1-pro, o3, o3-mini, o3-mini-high, o3-pro, o3-deep-research, o4-mini, o4-mini-high, o4-mini-deep-research
- **OpenAI (Open Source)**: openai/gpt-oss-20b, openai/gpt-oss-120b
- **Anthropic**: claude-opus-4-6, claude-opus-4-5, claude-opus-4-1, claude-sonnet-4-5, claude-sonnet-4, claude-haiku-4-5
- **Anthropic (Legacy)**: claude-3-5-sonnet-20241022, claude-3-5-haiku-20241022, claude-3-opus-20240229, claude-3-sonnet-20240229, claude-3-haiku-20240307
- **Gemini**: gemini-2.5-pro, gemini-2.5-flash, gemini-2.0-flash, gemini-2.0-flash-lite, gemini-1.5-pro (deprecated), gemini-1.5-flash (deprecated)
- **Gemini (preview, unreliable function calling)**: gemini-3-pro-preview, gemini-3-flash-preview, gemini-2.5-flash-lite
- **Together AI (Llama)**: Llama-4-Scout/Maverick (with function calling), Llama-3.x series (generation only)
- **Together AI (Qwen)**: Qwen2.5-72B/7B-Instruct-Turbo, Qwen2.5-Coder-32B, Qwen2.5-VL-72B, Qwen3-235B series, Qwen3-Coder-480B, Qwen3-Coder-Next, Qwen3-Next-80B, Qwen3-VL-32B/8B, QwQ-32B (with function calling)
- **Together AI (DeepSeek)**: DeepSeek-R1, DeepSeek-V3-0324, DeepSeek-V3.1 (with function calling)
- **Together AI (Kimi/Moonshot)**: Kimi-K2-Instruct, Kimi-K2-Instruct-0905, Kimi-K2-Thinking, Kimi-K2.5
- **Together AI (Mistral)**: Ministral-3-14B, Mistral-7B-v0.2/v0.3, Mistral-Small-24B
- **Together AI (Other)**: GLM-5.0/4.5-Air/4.7, Nemotron-Nano-9B, Arcee (Coder-Large, Maestro-Reasoning, Virtuoso-Large, trinity-mini), DeepCogito (cogito-v2 series), google/gemma-2b/3n, Refuel-LLM-2/2-Small, essentialai/rnj-1, marin-community/marin-8b
- **OpenRouter**: Access to 400+ models from 60+ providers via unified API:
- OpenAI (gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4.1, gpt-4o variants, gpt-5/5.1/5.2 and codex variants, o1, o3, o3-pro, o4-mini, codex-mini, gpt-oss, gpt-audio)
- Anthropic (claude-3-haiku, claude-3.5-haiku/sonnet, claude-3.7-sonnet, claude-sonnet-4/4.5, claude-haiku-4.5, claude-opus-4/4.1/4.5/4.6)
- Google (gemini-2.0-flash, gemini-2.5-flash/pro, gemini-3-flash/pro-preview, gemma-2-9b/27b, gemma-3-4b/12b/27b, gemma-3n-e4b)
- Meta Llama (llama-3-8b/70b, llama-3.1-8b/70b/405b, llama-3.2-1b/3b/11b-vision, llama-3.3-70b, llama-4-maverick/scout, llama-guard-2/3/4)
- DeepSeek (deepseek-chat/v3/v3.1/v3.2/v3.2-speciale, deepseek-r1/r1-0528/r1-turbo, deepseek-r1-distill variants, deepseek-coder-v2, deepseek-prover-v2)
- Qwen (qwen-2.5-7b/72b, qwen-turbo/plus/max, qwen3-8b/14b/30b/32b/235b, qwen3-coder/coder-plus/coder-next/coder-flash/coder-30b, qwen3-vl variants, qwq-32b, qwen3-next-80b, qwen3-max)
- Amazon Nova (nova-micro/lite/pro, nova-2-lite, nova-premier)
- Cohere (command-r, command-r-plus, command-a, command-r7b)
- X.AI Grok (grok-3/3-mini/3-beta/3-mini-beta, grok-4/4-fast, grok-4.1-fast, grok-code-fast-1)
- MiniMax (minimax-01, minimax-m1, minimax-m2/m2.1/m2-her)
- ByteDance Seed (seed-1.6, seed-1.6-flash, seed-2.0, seed-2.0-thinking)
- MoonshotAI (kimi-k2, kimi-k2-thinking, kimi-k2.5, kimi-dev-72b)
- Mistral (codestral, devstral/devstral-medium/devstral-small, mistral-large/medium/small, mixtral-8x7b/8x22b, ministral-3b/8b/14b, pixtral, voxtral)
- NVIDIA (llama-3.1-nemotron-70b/ultra-253b, llama-3.3-nemotron-super-49b, nemotron-nano-9b-v2/12b-v2-vl, nemotron-3-nano-30b)
- Z.AI/GLM (glm-5, glm-4-32b, glm-4.5/4.5-air/4.5v, glm-4.6/4.6v, glm-4.7/4.7-flash)
- AllenAI (olmo-2/3-7b/32b-instruct/think, olmo-3.1-32b-instruct/think, molmo-2-8b)
- Perplexity (sonar, sonar-pro, sonar-pro-search, sonar-deep-research, sonar-reasoning-pro)
- NousResearch (hermes-2-pro/3/4-llama series, hermes-4-70b/405b, deephermes-3)
- Baidu ERNIE (ernie-4.5 series including VL and thinking variants)
- And 30+ more providers (ai21, aion-labs, alfredpros, alpindale, anthracite-org, arcee-ai, bytedance, deepcogito, essentialai, ibm-granite, inception, inflection, kwaipilot, liquid, meituan, morph, nex-agi, opengvlab, prime-intellect, relace, sao10k, stepfun-ai, tencent, thedrummer, tngtech, upstage, writer, xiaomi, etc.)
**Embedding Models** (for RAG and semantic search):
- **OpenAI**: text-embedding-3-small, text-embedding-3-large, text-embedding-ada-002
- **Google**: text-embedding-004, gemini-embedding-001
- **Together AI**: BAAI/bge-large-en-v1.5, BAAI/bge-base-en-v1.5, m2-bert-80M-32k-retrieval, multilingual-e5-large-instruct, gte-modernbert-base
Each model in `MODEL_INFO` includes capability flags:
- `is_function_calling_supported`: Whether the model reliably supports tool/function calling
- `is_generation_supported`: Whether the model supports text generation
- `is_embedding_supported`: Whether the model is an embedding model
> **Note**: Additional models can be used, but context length, pricing, and capability information must be added to `src/kiss/core/models/model_info.py` for accurate token tracking, budget monitoring, and test filtering.
Token counts are extracted directly from API responses, ensuring accuracy and supporting multiple agents sharing the same model instance.
### Embedding Support
The framework provides embedding generation capabilities through the `get_embedding()` method on model instances:
- **OpenAI Models**: Full embedding support via OpenAI's embeddings API
- Default model: `text-embedding-3-small` (can be customized)
- Usage: `model.get_embedding(text, embedding_model="text-embedding-3-small")`
- **Together AI Models**: Full embedding support via Together AI's embeddings API
- Default model: `togethercomputer/m2-bert-80M-32k-retrieval` (can be customized)
- Usage: `model.get_embedding(text, embedding_model="togethercomputer/m2-bert-80M-32k-retrieval")`
- **Gemini Models**: Full embedding support via Google's embedding API
- Default model: `text-embedding-004` (can be customized; `gemini-embedding-001` also available)
- Usage: `model.get_embedding(text, embedding_model="text-embedding-004")`
- **Anthropic Models**: Embeddings not supported (raises `NotImplementedError`)
Embeddings are primarily used by the `SimpleRAG` system for document retrieval. When using `SimpleRAG`, ensure you use an OpenAI, Together AI, or Gemini model that supports embeddings.
## 🤗 Contributing
Contributions are welcome! Please ensure your code:
- Follows the KISS principle
- Passes all tests (`uv run pytest`)
- Passes linting (`uv run ruff check src/`)
- Passes type checking (`uv run mypy src/`)
- Passes type checking (`uv run pyright src/`)
## 📄 License
Apache-2.0
## ✍️ Authors
- Koushik Sen (ksen@berkeley.edu) | [LinkedIn](https://www.linkedin.com/in/koushik-sen-80b99a/) | [X @koushik77](https://x.com/koushik77)