Frameworks

Auto-instrumentation for agent frameworks.

Risicare provides deep integrations with popular agent frameworks.

Supported Frameworks

JavaScript SDK

The JS SDK supports 4 of the 10 frameworks above. The remaining 6 have no JavaScript integration in the Risicare SDK.

Framework	JS Import	Function
LangChain	`risicare/langchain`	`RisicareCallbackHandler` + `withSuppression()`
LangGraph	`risicare/langgraph`	`instrumentLangGraph()`
Instructor	`risicare/instructor`	`patchInstructor()`
LlamaIndex	`risicare/llamaindex`	`RisicareLlamaIndexHandler`

What's Captured

Framework integrations provide rich observability beyond basic LLM tracing:

Feature	Description
Agent Identity	Name, role, and hierarchy
Decision Flow	Think/Decide/Act phases
Tool Execution	Tool calls with inputs/outputs
Graph State	The names of the graph state keys, and one span for each node in step order (LangGraph). With content capture on, each node span also carries its input state and its update as text
Inter-Agent Messages	Communication between agents
Iteration Tracking	Loop counts and convergence

Provider Span Behavior

Some frameworks suppress the provider span of a call that the framework itself makes, so the call is counted once. Others let provider spans appear alongside framework spans. A provider call that your own code makes outside a framework call keeps its span. LiteLLM, DSPy and LlamaIndex suppress every provider call made while their call runs (LiteLLM: on the same thread), not only their own.

Framework	Provider Suppression	Notes
LangChain	Yes	The provider call of a model run gets no second span
LiteLLM	Yes	Suppressed from the pre-call hook to the end of the call
DSPy	Yes	Suppresses for LM calls
LlamaIndex	Selective	Non-streaming LLM and embedding calls only; a streamed LLM call also gets the provider span
LangGraph	No	A node that calls a LangChain model gets no provider span (the LangChain rule); a direct provider call does
CrewAI	No	Provider spans appear as children
AutoGen	No	Provider spans appear as children
OpenAI Agents	No	Provider spans appear as children when the Agents SDK uses Chat Completions; by default it uses the Responses API, which is not traced
Instructor	No	Provider spans appear as children
Pydantic AI	No	Provider spans appear as children, for models whose API the SDK traces (OpenAI: Chat Completions)

Auto vs Manual Instrumentation

Auto-Instrumentation

Call risicare.init(), or use Tier 0 (import risicare with RISICARE_API_KEY and RISICARE_TRACING=true):

import risicare
risicare.init()
 
# Framework code is automatically traced
from langgraph.graph import StateGraph
# ... your agent code

Manual Instrumentation

Add explicit context for more control:

from risicare import agent, trace_think, trace_decide, trace_act
 
@agent(name="researcher", role="specialist")
def research_agent(query: str):
    @trace_think
    def analyze():
        return analyze_query(query)
 
    @trace_decide
    def plan():
        return create_plan(analysis)
 
    @trace_act
    def execute():
        return run_search(plan)
 
    analysis = analyze()
    plan = plan()
    return execute()

Framework Detection

Risicare automatically detects which frameworks are installed and instruments them:

import risicare
 
# Returns a dict keyed by module name; each value describes that
# module's instrumentation state
print(risicare.get_instrumented_modules())