CrewAI
Auto-instrumentation for CrewAI multi-agent crews.
Risicare provides deep integration with CrewAI for multi-agent crew observability.
Python only
Version Compatibility
crewai >= 0.50.0.Installation
pip install 'risicare[crewai]'
# or
pip install risicare crewaiBasic Usage
import risicare
from crewai import Agent, Task, Crew
risicare.init()
# Define agents as usual - they're automatically traced
researcher = Agent(
role="Researcher",
goal="Find accurate information",
backstory="Expert at research"
)
writer = Agent(
role="Writer",
goal="Write compelling content",
backstory="Skilled technical writer"
)
# Create tasks
research_task = Task(
description="Research the topic",
agent=researcher
)
write_task = Task(
description="Write the article",
agent=writer
)
# Run crew - fully traced
crew = Crew(agents=[researcher, writer], tasks=[research_task, write_task])
result = crew.kickoff()What's Captured
Agent Details
| Field | Description |
|---|---|
agent.role | "orchestrator" (crew kickoff) or "worker" (agent execution) |
agent.type | "crewai" |
agent.name | The agent's role (or its name when it has no role), on worker spans. The worker span is crewai.agent/{role} |
Crew & Task Execution
| Field | Description |
|---|---|
framework.crewai.agent_count | Number of agents in the crew |
framework.crewai.agent_roles | List of agent roles (up to 10), on the span of a synchronous kickoff. kickoff_async sends two nested crewai.crew spans on current CrewAI; the inner one has the roles |
framework.crewai.agent_goal | The agent's goal (when content tracing is enabled) |
framework.crewai.agent_output | Worker agent output (when content tracing is enabled) |
framework.crewai.result | Final crew result (when content tracing is enabled) |
framework.crewai.task_count | Number of tasks |
framework.crewai.task_description | Task description, cut to 100 characters, when content tracing is enabled. Without it, the value is <risicare:content-omitted> |
Agent Hierarchy
Hierarchical crews are reflected in the span tree:
crew = Crew(
agents=[manager, worker1, worker2],
tasks=[task],
process=Process.hierarchical,
manager_llm=ChatOpenAI(model="gpt-4o")
)The span hierarchy shows which agents executed and in what order. Manager and worker relationships are visible through the parent-child span structure.
Tool Usage
A tool that a CrewAI agent calls runs inside that agent's crewai.agent/{name} span, so its work is attributed to the right worker:
from crewai_tools import SerperDevTool
search_tool = SerperDevTool()
researcher = Agent(
role="Researcher",
tools=[search_tool]
)Per-tool spans are not emitted
The CrewAI integration instruments crew kickoff and agent execution — it does not emit a separate span per tool call. If a tool makes an LLM call, that call is captured by provider instrumentation as a child of the agent span; a non-LLM tool (e.g. a web-search API) shows up as elapsed time inside the agent span rather than as its own crewai.tool/* span. Dedicated tool/task/delegation spans are on the roadmap.
Tasks
Task information is surfaced as attributes on the spans above, not as separate task spans: framework.crewai.task_count on the crew span and framework.crewai.task_description on the worker (crewai.agent) span. Dependencies you declare with context=[...] shape how CrewAI orders execution, but they are not emitted as separate dependency spans.
research_task = Task(
description="Research the topic",
agent=researcher
)
write_task = Task(
description="Write based on research",
agent=writer,
context=[research_task] # influences execution order; not a separate span
)Provider Spans
CrewAI instrumentation creates agent/framework-level spans. Underlying LLM calls (e.g., OpenAI, Anthropic) are traced separately by provider instrumentation, giving you both framework-level and LLM-level visibility.
Visualization
View crew execution in the dashboard:
- The trace page: the span waterfall and the timeline of the run
- The Agents page: statistics for each agent