Instrument

The Instrument section covers everything you need to add observability to your AI agents.

Overview

Risicare provides multiple ways to instrument your code:

1. Auto-Instrumentation - Zero-code tracing of LLM providers
2. SDK Decorators - Rich context with @agent, @trace_* decorators
3. Framework Integrations - Native support for LangGraph, CrewAI, AutoGen

Choose Your Approach

Progressive Integration

Start simple and add depth as needed:

export RISICARE_API_KEY=rsk-your-api-key
export RISICARE_TRACING=true

import risicare  # the one line you add
# ... your agent code, unchanged ...

import risicare
risicare.init(environment="production")

from risicare import agent
 
@agent(name="planner", role="orchestrator")
def plan(objective):
    pass

from risicare import session_context
 
with session_context(session_id=user_session):
    agent.run(query)

from risicare import trace_think, trace_decide, trace_act
 
@trace_think
def analyze(): pass
 
@trace_decide
def choose(): pass
 
@trace_act
def execute(): pass

from risicare import trace_message, trace_delegate
 
@trace_message
def send_to_reviewer(msg): pass
 
@trace_delegate
def assign_subtask(task): pass

What Gets Captured

Prompt and completion text is recorded only when two switches are on: trace_content=True in the SDK (off by default), and the project's content setting. In the JavaScript SDK, patchOpenAI and patchAnthropic record the text; a streamed call records the prompt only. The other JavaScript providers record no text.

The text of an error is not covered by these switches. The exception message, the stack trace, the span status message and the error.message attribute are exported also when content capture is off. If a provider's error message repeats part of a prompt, that text is in those fields. To filter them in Python, use mask: it receives statusMessage, exceptions.{i}.message, exceptions.{i}.stacktrace and error.message.

Next Steps