| CrewAI | Swarms API |
|---|---|
Agent(role, goal, backstory, llm) | agent with agent_name, system_prompt, model_name |
Task(description, agent, expected_output) | Top-level task string; role in system_prompt |
Crew(agents, tasks, process=sequential) | SequentialWorkflow via /v1/swarm/completions |
Crew(agents, tasks, process=hierarchical) | HierarchicalSwarm via /v1/swarm/completions |
crew.kickoff() | POST /v1/swarm/completions |
crew.kickoff_async() | Same endpoint; async via httpx.AsyncClient |
@tool / BaseTool | tools_list_dictionary (OpenAI-style function schemas) on agent spec |
manager_llm | Director agent in HierarchicalSwarm |
memory=True | Stateless per-request; add external memory for cross-session |
verbose=True | Full agent outputs returned in outputs response field |
Side-by-Side: Sequential Crew
CrewAI
from crewai import Agent, Task, Crew, Process
from langchain_openai import ChatOpenAI
llm = ChatOpenAI(model="gpt-4.1")
researcher = Agent(
role="Senior Research Analyst",
goal="Uncover cutting-edge developments in AI",
backstory="You work at a leading tech think tank with 10 years of experience.",
llm=llm,
verbose=True,
)
writer = Agent(
role="Tech Content Strategist",
goal="Craft compelling content on tech advancements",
backstory="You are a renowned content strategist known for insightful articles.",
llm=llm,
verbose=True,
)
research_task = Task(
description="Conduct comprehensive research on the latest AI advancements in 2025.",
expected_output="A detailed report with key findings, trends, and implications.",
agent=researcher,
)
write_task = Task(
description="Write an engaging blog post based on the research findings.",
expected_output="A 500-word blog post formatted in markdown.",
agent=writer,
)
crew = Crew(
agents=[researcher, writer],
tasks=[research_task, write_task],
process=Process.sequential,
verbose=True,
)
result = crew.kickoff()
print(result.raw)
Swarms API
import os
import requests
result = requests.post(
"https://api.swarms.world/v1/swarm/completions",
headers={"x-api-key": os.environ["SWARMS_API_KEY"], "Content-Type": "application/json"},
json={
"name": "AI Research and Writing Crew",
"description": "Research AI advancements then write a blog post",
"swarm_type": "SequentialWorkflow",
"task": "Research the latest AI advancements in 2025 and write an engaging 500-word blog post.",
"agents": [
{
"agent_name": "Senior Research Analyst",
"system_prompt": (
"You are a Senior Research Analyst at a leading tech think tank with 10 years "
"of experience. Your goal is to uncover cutting-edge developments in AI. "
"Conduct comprehensive research and return a detailed report with key findings, "
"trends, and implications."
),
"model_name": "gpt-4.1",
"max_loops": 1,
"temperature": 0.3,
},
{
"agent_name": "Tech Content Strategist",
"system_prompt": (
"You are a renowned Tech Content Strategist known for insightful, engaging articles. "
"Your goal is to craft compelling content on tech advancements. "
"Using the research provided, write a 500-word blog post formatted in markdown."
),
"model_name": "gpt-4.1",
"max_loops": 1,
"temperature": 0.6,
},
],
"max_loops": 1,
},
timeout=120,
).json()
final_output = result["outputs"]
print(final_output)
Agent(role, goal, backstory)→ combine into a singlesystem_promptTask(description, expected_output)→ merged intosystem_promptas instructionsProcess.sequential→"swarm_type": "SequentialWorkflow"crew.kickoff()→POSTrequest
Side-by-Side: Hierarchical Crew
CrewAI
from crewai import Agent, Task, Crew, Process
from langchain_openai import ChatOpenAI
manager_llm = ChatOpenAI(model="gpt-4.1")
analyst = Agent(
role="Data Analyst",
goal="Analyze financial data and extract key metrics",
backstory="Expert in financial modeling and data interpretation.",
llm=ChatOpenAI(model="gpt-4.1"),
)
writer = Agent(
role="Report Writer",
goal="Write clear financial reports",
backstory="Specialist in translating complex data into readable reports.",
llm=ChatOpenAI(model="gpt-4.1"),
)
analysis_task = Task(
description="Analyze Q4 2024 earnings data and extract key performance indicators.",
expected_output="Structured KPI report with growth metrics.",
agent=analyst,
)
report_task = Task(
description="Write an executive summary report based on the analysis.",
expected_output="A 300-word executive summary.",
agent=writer,
)
crew = Crew(
agents=[analyst, writer],
tasks=[analysis_task, report_task],
process=Process.hierarchical,
manager_llm=manager_llm,
verbose=True,
)
result = crew.kickoff()
Swarms API
import os
import requests
result = requests.post(
"https://api.swarms.world/v1/swarm/completions",
headers={"x-api-key": os.environ["SWARMS_API_KEY"], "Content-Type": "application/json"},
json={
"name": "Financial Report Crew",
"description": "Hierarchical crew for financial analysis and report writing",
"swarm_type": "HierarchicalSwarm",
"task": "Analyze Q4 2024 earnings data, extract key KPIs, and write a 300-word executive summary.",
"agents": [
{
"agent_name": "Data Analyst",
"system_prompt": (
"You are an expert Data Analyst specializing in financial modeling. "
"Analyze the provided financial data and extract key performance indicators, "
"growth metrics, and trends. Return a structured KPI report."
),
"model_name": "gpt-4.1",
"max_loops": 1,
"temperature": 0.2,
},
{
"agent_name": "Report Writer",
"system_prompt": (
"You are a specialist Report Writer who translates complex financial data into "
"clear, readable reports. Using the analysis provided, write a 300-word "
"executive summary suitable for C-suite readers."
),
"model_name": "gpt-4.1",
"max_loops": 1,
"temperature": 0.4,
},
],
"max_loops": 1,
},
timeout=120,
).json()
print(result["outputs"])
Side-by-Side: Tool-Using Agent
CrewAI
from crewai import Agent, Task, Crew
from crewai_tools import SerperDevTool, WebsiteSearchTool
search_tool = SerperDevTool()
web_tool = WebsiteSearchTool()
researcher = Agent(
role="Market Researcher",
goal="Find latest market data on electric vehicles",
backstory="Expert market researcher with access to real-time data.",
tools=[search_tool, web_tool],
llm=ChatOpenAI(model="gpt-4.1"),
)
task = Task(
description="Research current EV market trends, top players, and growth forecasts.",
expected_output="A market research report with data and sources.",
agent=researcher,
)
crew = Crew(agents=[researcher], tasks=[task])
result = crew.kickoff()
Swarms API
import os
import requests
result = requests.post(
"https://api.swarms.world/v1/agent/completions",
headers={"x-api-key": os.environ["SWARMS_API_KEY"], "Content-Type": "application/json"},
json={
"task": "Research the current electric vehicle market trends, top players, and growth forecasts.",
"tools_enabled": ["auto_search"],
"agent_config": {
"agent_name": "Market Researcher",
"system_prompt": (
"You are an expert Market Researcher with access to real-time data. "
"Research current EV market trends, top players, and growth forecasts. "
"Return a comprehensive market research report with data and sources."
),
"model_name": "gpt-4.1",
"max_loops": 1,
"temperature": 0.3,
},
},
timeout=90,
).json()
print(result["outputs"])
Migrating Agent Definitions
The most mechanical part of a CrewAI migration is convertingAgent(role, goal, backstory) to a system_prompt.
CrewAI
Agent(
role="Senior Data Scientist",
goal="Build accurate predictive models",
backstory="PhD in statistics with 8 years of ML experience at Fortune 500 companies.",
llm=ChatOpenAI(model="gpt-4.1"),
verbose=True,
allow_delegation=False,
max_iter=3,
)
Swarms API
{
"agent_name": "Senior Data Scientist",
"system_prompt": (
"You are a Senior Data Scientist with a PhD in statistics and 8 years of ML "
"experience at Fortune 500 companies. Your goal is to build accurate predictive "
"models. Approach every problem methodically, show your reasoning, and always "
"validate your assumptions."
),
"model_name": "gpt-4.1",
"max_loops": 3, # replaces max_iter
"temperature": 0.2,
}
role+goal+backstory→system_prompt(combine into natural prose)max_iter→max_loopsallow_delegation=False→ not needed; agents don’t self-delegate unless you define edgesverbose=True→ always on; full outputs are in the response
Migrating Task Definitions
CrewAI
Task(
description=(
"Analyze the provided customer churn dataset. "
"Identify the top 5 factors driving churn and estimate their impact."
),
expected_output=(
"A structured report listing the top 5 churn factors with statistical evidence "
"and recommended retention strategies."
),
agent=analyst,
context=[data_prep_task],
)
Swarms API
# task description + expected_output merge into system_prompt.
# Context from prior tasks is handled automatically via sequential workflow ordering.
{
"agent_name": "Churn Analyst",
"system_prompt": (
"You are a customer analytics expert. Analyze the provided churn dataset. "
"Identify the top 5 factors driving churn with statistical evidence. "
"Return a structured report with each factor, its impact estimate, and "
"a recommended retention strategy."
),
"model_name": "gpt-4.1",
"max_loops": 1,
"temperature": 0.2,
}
# "task": "Analyze the following churn dataset: ..."
Async Kickoff
CrewAI
import asyncio
async def run():
result = await crew.kickoff_async()
print(result.raw)
asyncio.run(run())
Swarms API
import asyncio
import httpx
import os
async def run():
async with httpx.AsyncClient() as client:
response = await client.post(
"https://api.swarms.world/v1/swarm/completions",
headers={"x-api-key": os.environ["SWARMS_API_KEY"], "Content-Type": "application/json"},
json={ ... },
timeout=300,
)
result = response.json()
print(result["outputs"])
asyncio.run(run())
Key Differences to Keep in Mind
| Concern | CrewAI | Swarms API |
|---|---|---|
| Agent memory | memory=True per agent | Stateless; use external DB (Redis, Postgres) |
| Inter-task context | context=[task_a, task_b] | Automatic in sequential workflows; use edges in graph workflows |
| Human input | human_input=True on Task | Not currently supported in API |
| Output parsing | output_pydantic, output_json | Not a dedicated field; instruct the model to return structured JSON via system_prompt/task, or define a schema in tools_list_dictionary |
| Rate limiting | Depends on LLM provider | Managed by the API; see rate limits |
| Cost tracking | External | usage.token_cost in every response |