CrewAI

• Agent = role + goal + backstory + tools + LLM; these strings are compiled into the system prompt.
• Task = description + expected output + assigned agent, optionally with context from earlier tasks and output_pydantic for typed results.
• Crew = agents + tasks + process (sequential or hierarchical, where a manager LLM delegates).
• Flows (newer) add event-driven, @start/@listen/@router orchestration with state, sitting above crews for deterministic control flow.
• Memory modules (short-term, long-term, entity) are opt-in and backed by embeddings plus a local store.

• Demos and internal tools where a "researcher → writer → reviewer" pipeline maps cleanly onto roles.
• Content generation and report drafting with clear stage boundaries.
• Non-specialist teams that want a readable YAML/Python description of who does what.

• Hidden prompts: role/goal/backstory are wrapped in substantial framework prompt text; what the model actually sees is not what you wrote.
• The persona abstraction encourages multi-agent designs where one agent with a good prompt would be cheaper and more reliable (When Not to Use Multi-Agent).
• Token cost multiplies: each task re-sends context, and hierarchical mode adds manager round-trips.
• Failure handling is coarse — a mid-crew tool error or a hallucinated delegation is hard to intercept, retry, or checkpoint.
• Debuggability: verbose console logs are the primary tool; tracing hooks exist but are less mature than LangSmith or OpenTelemetry-first stacks.

• Production workloads with strict latency or cost budgets.
• Anything needing deterministic control flow with approvals — use a state graph.
• Single-agent problems; the role machinery adds prompts, tokens, and failure surface for nothing.

1from crewai import Agent, Task, Crew, Process
2from crewai_tools import SerperDevTool
3 
4search = SerperDevTool()
5 
6researcher = Agent(
7    role="Market researcher",
8    goal="Find current, sourced facts about {topic}",
9    backstory="Meticulous analyst who cites sources.",   # becomes prompt text
10    tools=[search],
11)
12writer = Agent(role="Technical writer", goal="Write a 300-word brief", backstory="Plain-English explainer.")
13 
14research = Task(description="Research {topic}", expected_output="Bullet list with URLs", agent=researcher)
15brief = Task(description="Write the brief from the research", expected_output="Markdown brief",
16             agent=writer, context=[research])
17 
18crew = Crew(agents=[researcher, writer], tasks=[research, brief], process=Process.sequential)
19result = crew.kickoff(inputs={"topic": "vector database pricing"})
20print(result.raw)