Sentragent

Test your AI agents before they break something in production.

Gartner predicts that over 40% of enterprise AI agent projects will fail by 2027 — not because of model limitations, but because of insufficient controls before and after deployment.

Sentragent is an open-source SDK for instrumenting, testing, and continuously monitoring AI agents in production. Framework-agnostic by design: Sentinel only needs a plain agent(prompt) -> string callable, so it works with LangChain, CrewAI, Mastra, the OpenAI Agents SDK, or a fully custom agent — no framework-specific integration package required. See examples/ for LangChain, Mastra, CrewAI, and the OpenAI Agents SDK.

Why Sentragent

• Automatic adversarial scenario generation from your existing prompts — no need to hand-write your test cases.
• Configurable behavioral scoring (LLM-as-judge) — your business rules, not generic ones.
• Framework-agnostic — one interface across LangChain, CrewAI, Mastra, the OpenAI Agents SDK, or anything you built yourself.

Production drift detection and a native GitHub Actions integration for running your scenarios on every deploy are on the roadmap — see CHANGELOG.md.

Installation

# Python
pip install sentragent

# TypeScript / Node
npm install sentragent

To install from source instead (e.g. to contribute):

# Python
git clone https://github.com/Sentragent/sentragent-sdk.git
cd sentragent-sdk/python
pip install -e .

# TypeScript / Node
git clone https://github.com/Sentragent/sentragent-sdk.git
cd sentragent-sdk/typescript
npm install && npm run build

Quickstart

Zero setup, using the built-in library of common failure-mode scenarios:

from sentragent import Sentinel

def my_agent(prompt: str) -> str:
    return call_your_agent(prompt)

sentinel = Sentinel(agent=my_agent)
report = sentinel.run_scenarios(auto_generate=5)
print(report.summary())

Scenarios generated dynamically from your agent's own system prompt — bring your own LLM call (OpenAI, Anthropic, whatever you already use):

sentinel = Sentinel(agent=my_agent)
report = sentinel.run_scenarios(
    system_prompt=my_agent_system_prompt,
    llm=my_llm_call,       # any Callable[[str], str]
    auto_generate=10,
)
print(report.summary())

import { Sentinel } from "sentragent";

const sentinel = new Sentinel({ agent: myAgent });

// Zero setup:
const report = await sentinel.runScenarios({ autoGenerate: 5 });

// Or generated from your agent's own system prompt:
const report2 = await sentinel.runScenarios({
  systemPrompt: myAgentSystemPrompt,
  llm: myLlmCall, // (prompt: string) => Promise<string> | string
  autoGenerate: 10,
});

console.log(report.summary());

Project status

Sentragent is in active development (design partner phase). The API is not yet stable. See CHANGELOG.md for version history.

Design Partner Program

Looking for 5-10 teams with at least one AI agent in pre-production or production to test the tool on their real agents. Free lifetime access to the Team tier, direct influence on the roadmap, lifetime preferred pricing.

Apply: calendly.com/marlinibukun/sentragent-design-partner-call

Repo structure

sentragent-sdk/
├── python/           # Python SDK (package "sentragent")
├── typescript/       # TypeScript/Node SDK (package "sentragent")
├── examples/         # Integration examples (LangChain, Mastra, CrewAI, OpenAI Agents SDK)
├── CHANGELOG.md
└── LICENSE

License

Apache License 2.0 — see LICENSE.

Contact

EMAIL_ADDRESS · built by jaceX10