sondera-harness 0.7.1

Sondera Harness SDK for Python - Agent governance and policy enforcement

Sondera Harness

Deterministic guardrails for AI agents.

Open-source. Works with LangGraph, ADK, Strands, or any custom agent.

Docs · Quickstart · Examples · Slack

---

What is Sondera Harness?

Sondera Harness evaluates Cedar policies before your agent's actions execute. When a policy denies an action, the agent gets a reason why and can try a different approach. Same input, same verdict. Deterministic, not probabilistic.

Example policy:

forbid(principal, action, resource)
when { context has parameters_json && context.parameters_json like "*rm -rf*" };

This policy stops your agent from running rm -rf, every time.

Quickstart

Try it now: - no install required.

1. Install

uv add "sondera-harness[langgraph]"   # or: pip install "sondera-harness[langgraph]"

Works with LangChain/LangGraph, Google ADK, Strands, and custom agents.

2. Add to Your Agent (LangGraph)

from langchain.agents import create_agent
from sondera.harness import SonderaRemoteHarness
from sondera.langgraph import SonderaHarnessMiddleware, Strategy, create_agent_from_langchain_tools

# Analyze your tools and create agent metadata
sondera_agent = create_agent_from_langchain_tools(
    tools=my_tools,
    agent_id="langchain-agent",
    agent_name="My LangChain Agent",
    agent_description="An agent that helps with tasks",
)

# Create harness with agent
harness = SonderaRemoteHarness(agent=sondera_agent)

# Create middleware
middleware = SonderaHarnessMiddleware(
    harness=harness,
    strategy=Strategy.BLOCK,  # or Strategy.STEER
)

# Create agent with middleware
agent = create_agent(
    model=my_model,
    tools=my_tools,
    middleware=[middleware],
)

[!NOTE] This example uses Sondera Platform (free account), which also enables the TUI below. For local-only development, see CedarPolicyHarness.

3. See It in Action

uv run sondera   # or: sondera (if installed via pip)

Why Sondera Harness?

• Steer, don't block: Denied actions include a reason. Return it to the model, and it tries something else.
• Deterministic: Stop debugging prompts. Rules are predictable.
• Drop-in integration: Native middleware for LangGraph, Google ADK, and Strands.
• Full observability: Every action, every decision, every reason. Audit-ready.

Documentation

• Quickstart
• Writing Policies
• Integrations
• Reference

Community

• Slack for questions and feedback
• GitHub Issues for bugs
• Contributing for development setup

License

MIT