Every agent action is scored against our cross-network behavioral baseline. See how your agent compares to hundreds of others. Block the dangerous ones before they execute. The independent safety layer that learns from every agent in the network.
{ "tool_name": "send_email", "decision": "block", "risk_score": 0.91, "percentile": 97, "comparison_group_size": 847, "reason": "97th percentile risk across 847 agent deployments" }
The Problem
LangChain decides which tool to call. CrewAI decides how to orchestrate. Sansin decides whether the action should happen at all.
How It Works
One function call. Three steps. Every agent action compared to hundreds of others.
Your agent
Your agent wants to send_email, delete_file, or query_database. Before it executes, the call goes through SansinGate.
from sansin import SansinGate, wrap_tools_with_gate gate = SansinGate(api_key="sk_a1b2c3...") tools = wrap_tools_with_gate([email_tool, db_tool], gate) agent = create_agent(tools=tools) # Every tool call now goes through Sansin
Sansin scores against the baseline
Sansin computes a risk score, then ranks it against our cross-network behavioral baseline. send_email to one person? 12th percentile, allow. send_email to 500 recipients? 97th percentile across 847 agent deployments, block.
The baseline learns from every agent in the network. Your percentile ranking gets more meaningful with every deployment. Thompson Sampling blends in your team's corrections after 50+ overrides.
Every decision is logged with percentile ranking and full reasoning. Override any decision, and the network learns from your correction.
Safe execution
Allowed actions execute normally. Blocked actions return a structured decision with risk_score, percentile ranking, comparison_group_size, and a recommendation. Your agent handles the response.
Fail-open by default so Sansin never breaks your agent. Configurable fail-closed for high-stakes environments.
Try It
Click an example to see the API response with percentile ranking.
Use Cases
LangChain, CrewAI, custom agents. Every tool call scored against our cross-network baseline with percentile ranking. The SDK is 4 lines of Python.
Should you send this notification? Sansin's original use case. Same decision engine, proven in production. 37% fewer notifications, same revenue.
Autonomous devices making real-world decisions. Same behavioral baseline, same percentile ranking. On the 2027 roadmap.
What We Believe
Every query is tenant-scoped. Every model is isolated. Agent data stays where it belongs. We designed for this from day one, not as an afterthought.
If the smartest thing to do is not take an action, we don't take it. The best decision engines know when to say no.
You can see why an action was allowed or blocked, what the risk score was, and how confident the model was. No black boxes.
Something went wrong. Try again or email us directly.