back
loading skill details...
CrewAI agent design and configuration. Use when creating, configuring, or debugging crewAI agents — choosing role/goal/backstory, selecting LLMs, assigning…
CrewAI Agent Design Guide
How to design effective agents with the right role, goal, backstory, tools, and configuration.
The 80/20 Rule
Spend 80% of your effort on task design, 20% on agent design. A well-designed task elevates even a simple agent. But even the best agent cannot rescue a vague, poorly scoped task. Get the task right first (see the design-task skill), then refine the agent.
0. How Many Agents Do You Actually Need?
Default to ONE agent. Add more only when the task genuinely splits into work that requires:
Different tools or permissions — e.g. one agent has Slack write access, another reads docs only.
Different personas the LLM must clearly switch between — a writer's voice is not a researcher's voice.
Different LLMs — a cheap model for mechanical steps, a stronger one for synthesis.
Different guardrails or output schemas — separate agents make the contract per stage explicit.
DO NOT add an agent just because the workflow has multiple steps. A single agent can:
Call multiple tools in sequence within one kickoff (search → scrape → summarize is one agent's loop).
Produce structured multi-section output in one response.
Iterate via its own tool-use loop without you orchestrating it as separate agents.
Cost calculus: every extra agent = at least one more LLM kickoff plus a context handoff. Splitting linear, single-persona work into multiple agents multiplies token cost and adds fragility for marginal quality wins.
Anti-pattern: Sequential mechanical steps as separate agents
❌ Three agents for what is one researcher's job:
source_finder = Agent(role="Finds URLs via Firecrawl search", tools=[firecrawl_search])
scraper = Agent(role="Scrapes URLs via Firecrawl scrape", tools=[firecrawl_scrape])
writer = Agent(role="Writes the report", ...)
✅ One researcher does the gathering loop; one writer synthesizes — two agents because the personas and LLMs genuinely differ:
researcher = Agent(role="Web Researcher", tools=[firecrawl_search, firecrawl_scrape], llm="anthropic/claude-haiku-4-5")
writer = Agent(role="Technical Report Writer", llm="anthropic/claude-sonnet-4-6")
The researcher's task description tells it to search, then scrape, then return structured findings. One LLM loop, multiple tool calls.
Anti-pattern: "Summarize then send" as two agents
❌ Two agents to read a string, summarize it, and post a Slack DM:
summarizer = Agent(role="Summarizer")
slack_messenger = Agent(role="Slack Sender", apps=["slack"])
✅ One agent with the connector and a task that tells it to summarize on top, then DM:
slack_dm_agent = Agent(
role="Slack Reporter",
goal="Post a Slack DM containing a one-paragraph summary plus the full markdown body.",
apps=["slack"],
)
# Task: "Read the report below. Write a 2-3 sentence executive summary at the top.
# Post a DM to {recipient_email} with the summary followed by the full body."
Heuristic
If two "agents" share the same persona, the same tool surface, and the same LLM, they are one agent with a longer task description.
Once you've decided "one agent is enough"
Use Agent.kickoff() directly inside a Flow method — no Crew, no Task ceremony. The Flow owns sequencing and state; each step is a single agent kickoff. See Section 4 — Agent.kickoff() — Direct Agent Execution below for the full pattern, and the upstream docs at https://docs.crewai.com/en/concepts/agents#direct-agent-interaction-with-kickoff.
Quick shape:
@listen(previous_step)
def my_step(self):
agent = Agent(role="…", goal="…", backstory="…", tools=[...])
result = agent.kickoff(
messages=f"Use this prior step's output: {self.state.prior_field}",
response_format=MyPydanticModel, # optional
)
self.state.my_field = result.pydantic # or result.raw
Reach for Crew.kickoff() only when a step genuinely benefits from multi-agent collaboration (delegation, hierarchical management, parallel specialists feeding one synthesis). For "one agent does one job", Agent.kickoff() inside a Flow listener is the right primitive.
Only after you've decided multi-agent is justified, read on for how to design each one.
1. The Role-Goal-Backstory Framework
Every agent needs three things: who it is, what it wants, and why it's qualified.
Role — Who the Agent Is
The role defines the agent's area of expertise. Be specific, not generic.
Bad
Good
Researcher
Senior Data Researcher specializing in {topic}
Writer
Technical Blog Writer for developer audiences
Analyst
Financial Risk Analyst with regulatory compliance expertise
The role directly shapes how the LLM reasons. A "Senior Data Researcher" will produce different output than a "Research Assistant" even with the same task.
Goal — What the Agent Wants
The goal is the agent's individual objective. It should be outcome-focused with quality standards.
Bad
Good
Do research
Uncover cutting-edge developments in {topic} and identify the top 5 trends with supporting evidence
Write content
Produce publication-ready technical articles that explain complex topics clearly for non-technical readers
Analyze data
Deliver actionable risk assessments with confidence levels and recommended mitigations
Backstory — Why the Agent Is Qualified
The backstory establishes expertise, experience, values, and working style. It's the agent's "personality prompt."
backstory: >
You're a seasoned researcher with 15 years of experience in AI/ML.
You're known for your ability to find obscure but relevant papers
and synthesize complex findings into clear, actionable insights.
You always cite your sources and flag uncertainty explicitly.
What to include in a backstory:
Years/depth of experience
Specific domain knowledge
Working style and values (e.g., "always cites sources", "prefers concise output")
Quality standards the agent holds itself to
What NOT to include:
Implementation details (tools, models, config)
Task-specific instructions (those go in the task description)
Arbitrary personality traits that don't affect output quality
2. Agent Configuration Reference
Essential Parameters
Agent(
role="...", # Required: agent's expertise area
goal="...", # Required: what the agent aims to achieve
backstory="...", # Required: context and personality
llm="openai/gpt-4o", # Optional: defaults to OPENAI_MODEL_NAME env var or "gpt-4"
tools=[...], # Optional: list of tool instances
)
Execution Control
Agent(
...,
max_iter=25, # Max reasoning iterations per task (default: 25)
max_execution_time=300, # Timeout in seconds (default: None — no limit)
max_rpm=10, # Rate limit: max API calls per minute (default: None)
max_retry_limit=2, # Retries on error (default: 2)
verbose=True, # Show detailed execution logs (default: False)
)
Tuning max_iter:
Default 25 is generous — most tasks finish in 3-8 iterations
Lower to 10-15 to fail faster when tasks are well-defined
If agent consistently hits max_iter, the task is too vague (fix the task, not the limit)
Tool Configuration
from crewai_tools import SerperDevTool, ScrapeWebsiteTool, FileReadTool
Agent(
...,
tools=[SerperDevTool(), ScrapeWebsiteTool()], # Agent-level tools
)
Key rules:
An agent with no tools will hallucinate data when asked to search, fetch, or read files — always provide tools for tasks that require external data
Prefer fewer, focused tools over many tools — too many tools confuses the agent
Tools can also be assigned at the task level for task-specific access (see design-task skill)
Agent-level tools are available for all tasks the agent performs; task-level tools override for that specific task
LLM Selection
Agent(
...,
llm="openai/gpt-4o", # Main reasoning model
function_calling_llm="openai/gpt-4o-mini", # Cheaper model for tool calls only
)
Use function_calling_llm to save costs: the main llm handles reasoning while a cheaper model handles tool-calling mechanics.
Collaboration
Agent(
...,
allow_delegation=False, # Default: False — agent works alone
)
Set allow_delegation=True only when:
The agent is part of a crew with other specialized agents
The task genuinely benefits from the agent handing off subtasks
You're using hierarchical process where the manager delegates
Warning: Delegation without clear task boundaries leads to infinite loops or wasted iterations.
Planning (Plan-and-Execute Mode)
When a PlanningConfig is set on an agent, Agent.kickoff() (and Agent.execute_task()) routes through the new crewai.experimental.AgentExecutor. Instead of a single ReAct-style loop, the agent:
Generates a plan — a list of PlanSteps, each with a description and optional tool_to_use. Stored as state.todos.
Executes each step via a StepExecutor in an isolated multi-turn LLM loop (capped by max_step_iterations).
Observes the result via a PlannerObserver after every step — did the step succeed? Is the remaining plan still valid?
Routes the next action based on the agent's reasoning_effort setting (see below).
The presence of a PlanningConfig enables the mode. To disable: don't pass one, or set planning=False.
from crewai import Agent
from crewai.agent.planning_config import PlanningConfig
agent = Agent(
role="…",
goal="…",
backstory="…",
tools=[...],
planning_config=PlanningConfig(reasoning_effort="medium"), # most common
)
reasoning_effort — pick one
Level
After each step the planner...
Pick when
"low"
observes (validates success), marks the todo complete, continues. No replan, no refine.
You want plan visibility (todos, observations) but trust the agent to follow it linearly. Fastest.
"medium" (default)
observes; replans on failure only. Successful steps just continue.
The agent's tools can fail (network, exec, scrape) and you want graceful recovery without paying refinement cost on every success. The right default for sandbox-coding, research, and other tool-heavy loops.
"high"
observes, then routes through decide_next_action which can trigger early goal achievement, full replan, or lightweight refinement after every step.
The task changes shape based on intermediate findings, or you need maximum adaptiveness. Most LLM calls per run.
Source: crewai/experimental/agent_executor.py:450 (observe_step_result router) and crewai/agent/planning_config.py.
Other PlanningConfig knobs
PlanningConfig(
reasoning_effort="medium",
max_steps=20, # cap on planned steps (default 20)
max_replans=3, # max full re-plans before finalizing (default 3)
max_attempts=None, # planning refinement attempts during plan generation
max_step_iterations=15, # max LLM turns per step's StepExecutor (default 15)
step_timeout=None, # wall-clock seconds per step; None = no cap
system_prompt=None, # custom planning system prompt (uses default if None)
plan_prompt=None, # custom initial-plan prompt; placeholders: {description}, {expected_output}, {tools}, {max_steps}
refine_prompt=None, # custom refinement prompt
llm=None, # separate LLM for planning (else uses agent.llm)
)
Use llm="anthropic/claude-haiku-4-5" (cheap) for the planner while keeping agent.llm="anthropic/claude-opus-4-7" (strong) for execution — common cost optimization.
When to enable
Enable for autonomous loops where the agent picks its own steps and you want failure recovery (e.g. coding agent that writes → runs → patches; research agent that searches → scrapes → revises).
Skip for single-tool, single-purpose calls (e.g. "summarize this string", "post this Slack DM") — observation overhead doesn't pay off.
Cost shape
Every step gets a PlannerObserver LLM call (~1 extra call per step). On "medium" a failed step adds a replan call. On "high" every step adds a decide_next_action call too. For an N-step plan, expect roughly:
low: N execution + N observation = 2N calls
medium: 2N + (failures × 1 replan)
high: ~3N + replans/refines
Material at scale — measure before defaulting high for everything.
Custom plan_prompt
If you supply plan_prompt, include the placeholders the planner template expects: {description}, {expected_output}, {tools}, {max_steps}. The planner LLM gets these interpolated. Keep custom prompts focused on project-specific rules; let description/tools (auto-injected) carry the dynamic content.
Code Execution
Agent(
...,
allow_code_execution=True, # Enable code execution (default: False)
code_execution_mode="safe", # "safe" (Docker) or "unsafe" (direct) — default: "safe"
)
"safe" requires Docker installed and running — executes in a container
"unsafe" runs code directly on the host — only use in controlled environments
Context Window Management
Agent(
...,
respect_context_window=True, # Auto-summarize to stay within limits (default: True)
)
When True, the agent automatically summarizes prior context if it approaches the LLM's token limit. When False, execution stops with an error on overflow.
Date Injection
Agent(
...,
inject_date=True, # Add current date to task context (default: False)
date_format="%Y-%m-%d", # Date format (default: "%Y-%m-%d")
)
Enable for time-sensitive tasks (research, news analysis, scheduling).
Agent Guardrails
def validate_no_pii(result) -> tuple[bool, Any]:
"""Reject output containing PII."""
if contains_pii(result.raw):
return (False, "Output contains PII. Remove all personal information and try again.")
return (True, result)
Agent(
...,
guardrail=validate_no_pii,
guardrail_max_retries=3, # default: 3
)
Agent guardrails validate every output the agent produces. The agent retries on failure up to guardrail_max_retries.
Knowledge Sources
from crewai.knowledge.source.text_file_knowledge_source import TextFileKnowledgeSource
Agent(
...,
knowledge_sources=[
TextFileKnowledgeSource(file_paths=["company_handbook.txt"]),
],
embedder={
"provider": "openai",
"config": {"model": "text-embedding-3-small"},
},
)
Knowledge sources give agents access to domain-specific data via RAG. Use when agents need to reference large documents, policies, or datasets.
3. YAML Configuration (Recommended)
Define agents in agents.yaml for clean separation of config and code:
researcher:
role: >
{topic} Senior Data Researcher
goal: >
Uncover cutting-edge developments in {topic}
with supporting evidence and source citations
backstory: >
You're a seasoned researcher with 15 years of experience.
Known for finding obscure but relevant sources and
synthesizing complex findings into clear insights.
You always cite your sources and flag uncertainty.
# Optional overrides (uncomment as needed):
# llm: openai/gpt-4o
# max_iter: 15
# max_rpm: 10
# allow_delegation: false
# verbose: true
Then wire in crew.py:
@CrewBase
class MyCrew:
agents_config = "config/agents.yaml"
tasks_config = "config/tasks.yaml"
@agent
def researcher(self) -> Agent:
return Agent(
config=self.agents_config["researcher"],
tools=[SerperDevTool()],
)
Critical: The method name (def researcher) must match the YAML key (researcher:). Mismatch causes KeyError.
4. Agent.kickoff() — Direct Agent Execution
Use Agent.kickoff() when you need one agent with tools and reasoning, without crew overhead. This is the most common pattern in Flows.
Basic Usage
from crewai import Agent
from crewai_tools import SerperDevTool
researcher = Agent(
role="Senior Research Analyst",
goal="Find comprehensive, factual information with source citations",
backstory="Expert researcher known for thorough, evidence-based analysis.",
tools=[SerperDevTool()],
llm="openai/gpt-4o",
)
# Pass a string prompt — the agent reasons, uses tools, and returns a result
result = researcher.kickoff("What are the latest developments in quantum computing?")
print(result.raw) # str — the agent's full response
print(result.usage_metrics) # token usage stats
With Structured Output
from pydantic import BaseModel
class ResearchFindings(BaseModel):
key_trends: list[str]
sources: list[str]
confidence: float
result = researcher.kickoff(
"Research the latest AI agent frameworks",
response_format=ResearchFindings,
)
# Access via .pydantic (NOT directly — Agent.kickoff wraps the result)
print(result.pydantic.key_trends) # list[str]
print(result.pydantic.confidence) # float
print(result.raw) # raw string version
Note: Agent.kickoff() returns LiteAgentOutput — access structured output via result.pydantic. This differs from LLM.call() which returns the Pydantic object directly.
With File Inputs
result = researcher.kickoff(
"Analyze this document and summarize the key findings",
input_files={"document": FileInput(path="report.pdf")},
)
Async Variant
result = await researcher.kickoff_async(
"Research quantum computing breakthroughs",
response_format=ResearchFindings,
)
Agent.kickoff() in Flows (Recommended Pattern)
The most powerful pattern is orchestrating multiple Agent.kickoff() calls inside a Flow. The Flow handles state and sequencing; each agent handles its specific step:
from crewai import Agent
from crewai.flow.flow import Flow, listen, start
from crewai_tools import SerperDevTool, ScrapeWebsiteTool
from pydantic import BaseModel
class ResearchState(BaseModel):
topic: str = ""
research: str = ""
analysis: str = ""
report: str = ""
class ResearchFlow(Flow[ResearchState]):
@start()
def gather_data(self):
researcher = Agent(
role="Senior Researcher",
goal="Find comprehensive data with sources",
backstory="Expert at finding and validating information.",
tools=[SerperDevTool(), ScrapeWebsiteTool()],
)
result = researcher.kickoff(f"Research: {self.state.topic}")
self.state.research = result.raw
@listen(gather_data)
def analyze(self):
analyst = Agent(
role="Data Analyst",
goal="Extract actionable insights from raw research",
backstory="Skilled at pattern recognition and synthesis.",
)
result = analyst.kickoff(
f"Analyze this research and extract key insights:\n\n{self.state.research}"
)
self.state.analysis = result.raw
@listen(analyze)
def write_report(self):
writer = Agent(
role="Report Writer",
goal="Create clear, well-structured reports",
backstory="Technical writer who makes complex topics accessible.",
)
result = writer.kickoff(
f"Write a comprehensive report from this analysis:\n\n{self.state.analysis}"
)
self.state.report = result.raw
flow = ResearchFlow()
flow.kickoff(inputs={"topic": "AI agents"})
print(flow.state.report)
When to use Agent.kickoff() vs Crew.kickoff():
Use Agent.kickoff() when each step is a distinct agent and the Flow controls sequencing
Use Crew.kickoff() when multiple agents need to collaborate on related tasks within a single step
5. Specialist vs Generalist Agents
Note: Apply this section after you've decided you genuinely need multiple agents (see Section 0). If you only need one agent, "specialist vs generalist" is not the question — the question is just how to design that one agent.
When you do need multiple agents, prefer specialists. An agent that does one thing well outperforms one that does many things acceptably.
When to Use a Specialist
Task requires deep domain knowledge
Output quality matters more than speed
The task is complex enough to benefit from focused expertise
When a Generalist Is Acceptable
Simple tasks with clear instructions
Prototyping where you'll specialize later
Tasks that truly span multiple domains equally
Specialist Design Pattern
Instead of one "Content Writer" agent, create:
technical_writer — deep technical accuracy, code examples
copywriter — persuasive, audience-focused marketing copy
editor — grammar, consistency, style guide enforcement
Each specialist has a narrow role, specific goal, and backstory that reinforces their expertise.
6. Agent Interaction Patterns
Sequential (Default)
Agents work one after another. Each agent receives prior agents' outputs as context.
Researcher → Writer → Editor
Best for: linear pipelines where each step builds on the last.
Hierarchical
A manager agent delegates and validates. Task assignment is dynamic.
Crew(
agents=[researcher, writer, editor],
tasks=[research_task, writing_task, editing_task],
process=Process.hierarchical,
manager_llm="openai/gpt-4o",
)
Best for: complex workflows where task assignment depends on intermediate results.
Agent-to-Agent Delegation
When allow_delegation=True, an agent can ask another crew agent for help:
lead_researcher = Agent(
role="Lead Researcher",
goal="Coordinate research efforts",
backstory="...",
allow_delegation=True, # Can delegate to other agents in the crew
)
The agent will automatically discover other crew members and delegate subtasks as needed.
7. Common Agent Design Mistakes
Mistake
Impact
Fix
Generic role like "Assistant"
Agent produces unfocused, shallow output
Use specific expertise: "Senior Financial Analyst"
No tools for data-gathering tasks
Agent hallucinates data instead of searching
Always add tools when the task requires external info
Too many tools (10+)
Agent gets confused choosing between tools
Limit to 3-5 relevant tools per agent
Backstory full of task instructions
Agent mixes personality with task execution
Keep backstory about WHO the agent is; task details go in the task
allow_delegation=True by default
Agents waste iterations delegating trivially
Only enable when delegation genuinely helps
max_iter too high for simple tasks
Agent loops unnecessarily on vague tasks
Lower max_iter; fix the task description instead
No guardrail on critical output
Bad output passes through unchecked
Add guardrails for outputs that feed into production systems
Using expensive LLM for tool calls
Unnecessary cost for mechanical operations
Set function_calling_llm to a cheaper model
8. Agent Design Checklist
Before deploying an agent, verify:
Role is specific and domain-focused (not "Assistant" or "Helper")
Goal includes desired outcome AND quality standards
Backstory establishes expertise and working style
Tools are assigned for any task requiring external data
No excess tools — 3-5 per agent maximum
max_iter is tuned for expected task complexity (10-15 for simple, 20-25 for complex)
max_execution_time is set for production agents to prevent hangs
Guardrails are configured for critical outputs
LLM is appropriate for task complexity (don't use GPT-4 for classification)
Delegation is disabled unless genuinely needed
References
For deeper dives into specific topics, see:
Custom Tools — building your own tools with @tool decorator and BaseTool subclass
Memory & Knowledge — memory configuration, knowledge sources, embedder setup, scoping
For related skills:
getting-started — project scaffolding, choosing the right abstraction, Flow architecture
design-task — task description/expected_output best practices, guardrails, structured output, dependencies
ask-docs — query the live CrewAI documentation MCP server for questions not covered by these skillsdon't have the plugin yet? install it then click "run inline in claude" again.