Traigent

Traigent is an AI Agent infrastructure that allows companies to take AI agents out of the lab and deploy them at high scale with high confidence.

Our mission: Anything you can measure, we can improve. Whether it's accuracy, speed of response, cost, or any other business metric — we bring strong results that deliver real business value.

Runs multiple LLM trials — run python -m traigent.examples.quickstart for a no-cost demo, or call enable_mock_mode_for_quickstart() from traigent.testing in local tutorial code. Set TRAIGENT_RUN_COST_LIMIT=2.0 as a best-effort local guardrail and set hard caps with your LLM/cloud providers; actual billing is determined by those providers. The legacy TRAIGENT_MOCK_LLM=true env var remains for backwards compatibility and is disabled when ENVIRONMENT=production. See Cost Management.

Quick Install:

Install from PyPI — no clone required (Python 3.11+):

# pip
pip install "traigent[recommended]"

# uv (into the active venv, or `uv add` into a uv project)
uv pip install "traigent[recommended]"
uv add "traigent[recommended]"

python3 -m venv .venv
source .venv/bin/activate
pip install "traigent[recommended]"

python -m venv .venv
.venv\Scripts\Activate.ps1
pip install "traigent[recommended]"

For more options, see Installation details.

Try it now - no API keys needed:

pip install "traigent[integrations]"
python -m traigent.examples.quickstart

Or from a source checkout:

python hello_world.py

Here's what the quickstart does - one decorator, automatic optimization:

from langchain_openai import ChatOpenAI
import traigent

@traigent.optimize(
    configuration_space={
        "model": ["gpt-4o-mini", "gpt-4o"],
        "temperature": [0.0, 0.7, 1.0],
    },
    objectives=["accuracy"],
    eval_dataset="qa_samples.jsonl",
)
def answer(question: str) -> str:
    cfg = traigent.get_config()
    llm = ChatOpenAI(model=cfg["model"], temperature=cfg["temperature"])
    return llm.invoke(question).content

Using it in your own code

Add @traigent.optimize() to any function that calls an LLM — no framework required:

import traigent
import litellm                    # or openai, anthropic, requests …

@traigent.optimize(
    configuration_space={
        "model": ["gpt-4o-mini", "gpt-4o"],
        "temperature": [0.0, 0.7, 1.0],
    },
    objectives=["accuracy"],
    eval_dataset="path/to/your_evals.jsonl",
)
def your_function(question: str) -> str:
    cfg = traigent.get_config()
    response = litellm.completion(
        model=cfg["model"],
        temperature=cfg["temperature"],
        messages=[{"role": "user", "content": question}],
    )
    return response.choices[0].message.content


# Run it — this executes the trials and returns the winner
result = your_function.optimize_sync(max_trials=10)
print(result.best_config, result.best_score)

Works with any LLM provider — OpenAI, Anthropic, LiteLLM (100+ providers), or plain HTTP calls.

Portal · Quickstart · Examples · Skill · Walkthrough

Choose Your Path

🎬 See Traigent in Action — click to play demos

Demo
LLM Agent Optimization
Optimization Callbacks
Agent Configuration Hooks

🚀 Walkthrough — 8 runnable examples

The walkthrough examples use local mock mode through the quickstart/testing helpers with cost approval pre-set for dry runs — no provider API keys needed when calls go through LiteLLM or LangChain.

Browse reference examples → · Injection modes →

☁️ Traigent Portal Tracking

Connect to Traigent Portal to view results, compare trials, and collaborate. Portal tracking is enabled automatically when TRAIGENT_API_KEY is set; most users do not need any routing settings.

The default run uses Traigent's smart optimizer when portal credentials are available. Local grid and random searches still sync results to the portal when authenticated. Use offline=True only when a run must avoid Traigent backend egress; offline runs do not sync.

1. Sign up at portal.traigent.ai — verify your email to activate
2. Create an API key — click your name (top-right) → API Keys → + Create API Key
3. Connect — run traigent auth login or set export TRAIGENT_API_KEY="sk_..."
4. Run — portal tracking is automatic

@traigent.optimize(
    configuration_space={
        "model": ["gpt-4o-mini", "claude-haiku-4-5-20251001", "gemini-1.5-pro"],
        "temperature": [0.1, 0.5, 0.9],
    },
    objectives=["accuracy", "cost"],
    eval_dataset="data/qa_samples.jsonl",
)
def multi_provider_agent(question: str) -> str:
    config = traigent.get_config()
    response = litellm.completion(
        model=config.get("model"),
        temperature=config.get("temperature"),
        messages=[{"role": "user", "content": question}],
    )
    return response.choices[0].message.content

✨ Key Features

TraigentDemo — Streamlit playground, use cases, and research benchmarks

git clone https://github.com/Traigent/Traigent.git
cd Traigent
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[recommended]"

{"input": {"question": "What is AI?"}, "output": "Artificial Intelligence"}
{"input": {"question": "Explain ML"}, "output": "Machine learning uses data and algorithms"}

traigent optimize module.py -a grid -n 10   # Run optimization
traigent validate data.jsonl -o accuracy     # Validate dataset
traigent results list                        # List past runs
traigent plot <name> -p progress             # Visualize results
traigent auth login                          # Authenticate with portal
traigent --help                              # Full command reference

🛠️ Development

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[all,dev]"              # Install with dev dependencies
pytest                                   # Run tests
make format && make lint                 # Format and lint

Using uv (this repo commits uv.lock): run uv sync --all-extras --no-extra deepeval --dev, not bare uv sync --all-extras --dev. The deepeval extra is deliberately excluded from the all/recommended/enterprise bundles (opt-in only, pending an upstream telemetry-exfiltration fix — see pyproject.toml), but --all-extras installs every extra regardless, including the excluded one. The installed deepeval then imports langchain.schema, which the langchain version --all-extras also resolves no longer provides, and pytest tests/unit fails to collect. CI installs with pip install -e ".[all,dev]", which never selects deepeval, so this only shows up locally with uv sync --all-extras.

Architecture guide → · Project structure →

🤝 Contributing

We welcome bug reports and feature requests via GitHub Issues. For security vulnerabilities, please email security@traigent.ai.

📄 License

This project is dual-licensed: the GNU Affero General Public License v3.0 only (AGPL-3.0-only) - see LICENSE - or a Traigent commercial license under a separate written agreement (see COMMERCIAL-LICENSE.md). SPDX: AGPL-3.0-only OR LicenseRef-Traigent-Commercial. Commercial inquiries: legal@traigent.ai.

Get Started → | Examples → | Portal → | Skill → | Walkthrough → | GitHub Issues | Discussions