AgentTrust Docs

Overview

auto_instrument() monkey-patches popular AI SDKs at runtime to automatically send validations to AgentTrust — no decorator or client code required. Supported targets: OpenAI SDK, LangChain (LCEL + legacy), LangGraph.

Why It Matters

The lowest-friction integration for existing codebases. Add two lines at startup and all LLM calls are governed.

Prerequisites

pip install agentrust-py
# Plus your framework: openai, langchain, langgraph

Step-by-Step Guide

1. Call before framework imports (recommended)

from agentrust_sdk import auto_instrument
auto_instrument()  # must run before openai/langchain imports

import openai
# ... rest of app

2. Alternative: `install()` / `auto_install()`

A second, independent patcher lives in agentrust_sdk.auto. It covers Anthropic, which auto_instrument() does not, but omits LangGraph.

from agentrust_sdk.auto import install

install(agent_id="my-agent")   # patches OpenAI, Anthropic, LangChain callbacks

	`auto_instrument()`	`auto.install()`
OpenAI	✅	✅
LangChain	✅ LCEL + legacy	✅ callback handler
LangGraph	✅	❌
Anthropic	❌	✅
Reversible	✅ `remove_patches()`	❌

Prefer auto_instrument() unless you specifically need the Anthropic patch. For Anthropic with full control, ClaudeAgentGuard is the explicit alternative.

3. Disable auto-instrumentation

export AGENTRUST_AUTO_INSTRUMENT=false

4. What gets patched

Library	Patched methods
OpenAI	`chat.completions.create` — sync and async. The Responses API is not patched.
LangChain	`Runnable.invoke` (LCEL), plus legacy `BaseLLM.predict` and `Chain.__call__`
LangGraph	`CompiledStateGraph.invoke` and `.ainvoke`

Only the module-level OpenAI client is patched

auto_instrument() patches openai.chat.completions.create on the module-level default client and on the AsyncOpenAI class. Calls made through your own client = OpenAI() instance are not governed. Either use the module-level openai.chat.completions.create(...), or wrap your call site in @harness.

Patches are reversible: remove_patches() restores the original callables. A process restart is the safest rollback once traffic has been served.

Integration pattern A

From examples/README.md — Pattern A is auto-instrument for zero application code change.

Examples

from agentrust_sdk import auto_instrument, embed_gateway

embed_gateway()
auto_instrument(agent_id="my-openai-agent")

# Module-level call — this IS governed.
import openai
response = openai.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}],
)

# Instance call — this is NOT governed by auto_instrument().
# from openai import OpenAI
# client = OpenAI()
# client.chat.completions.create(...)   # ← bypasses the patch

See examples/minimal.py and examples/openai_direct.py.

Best Practices

Call auto_instrument() as early as possible in main() or __init__.py
Pass agent_id= to auto_instrument() — called directly it defaults to "auto-instrumented" and reads no env var. (AGENTRUST_AUTOLOAD_AGENT_ID is read only by the autoload bootstrap.)
Combine with embed_gateway() for local dev
Use @harness instead when you need explicit agent boundaries

Common Mistakes

Importing OpenAI before auto_instrument() (patch may not apply)
Expecting a meaningful agent_id for free — it defaults to "auto-instrumented"
Expecting your own OpenAI() instance to be governed — only the module-level client is patched
Using auto-instrument with Team-tier graph adapters (use AgentTrustNode instead)

Troubleshooting

Issue	Fix
Calls not validated	Confirm `AGENTRUST_ENABLED=true` and `AGENTRUST_AUTO_INSTRUMENT` is not `false`; check import order; confirm you are not calling through your own `OpenAI()` instance
Double validation	Don't combine auto-instrument + `@harness` on same path
LangGraph not patched	Ensure langgraph is installed; check logs

Auto-Instrumentation Overview