CrewAI is built around four core primitives: Agent (the team member), Task (the work to be done), Tool (capabilities agents can use), and Crew (the team that works together).

Specific, detailed backstories with concrete experience (years, brands, achievements) give the LLM rich context that shapes writing style, expertise level, and decision-making. Vague descriptions produce vague outputs.

expected_output defines what 'done' looks like - the format, structure, and content requirements. Without it, agents may produce outputs that don't meet your needs.

The docstring is critical - agents use it to understand when and how to use the tool. Without a clear docstring, agents won't know the tool's purpose or correct parameter values.

Knee-jerk price matching can trigger price wars and margin erosion. The crew should analyze whether it's a temporary promo, assess strategic implications, and provide recommendations with risk analysis.

While all agents should be accurate, the Quality Reviewer's specific job is to verify claims against source data. This is why multi-agent workflows with dedicated review steps catch errors that single-agent approaches miss.

Use a single crew instance and loop through products. Running all in one task would overload context; separate crews waste resources; one at a time works but controlled batching is more efficient.

Safety issues require immediate human escalation. Automated responses should express concern and request direct contact. Never offer product exchange for safety issues (liability concerns). Always escalate to Safety Team.

Instagram (visual, casual), Email (personalized, detailed), and Google Ads (concise, keyword-focused) have fundamentally different requirements. Specialized agents understand each platform's unique language and constraints.

Hierarchical is ideal when a manager needs to assess the situation and dynamically delegate to the right specialists. Sequential is better for predictable pipelines where each step clearly follows the previous.

Competitor weaknesses are your opportunities. If their customers have unmet needs (wide fit), you can capture that market segment through targeted marketing and potentially product development.

The context parameter creates task dependencies. When Task B has context=[Task A], it receives Task A's output as additional context, enabling sequential workflows where each step builds on previous work.

Human-in-the-loop patterns are essential for high-stakes decisions. The crew provides intelligent recommendations and handles routine changes, while significant changes get human oversight before execution.

Tools should return strings, typically JSON-formatted, because LLMs process text. JSON provides structure that agents can easily parse while remaining human-readable for debugging.

A hierarchical manager can analyze the specific issue and route to only the needed specialists. Sequential would waste resources running all agents for every issue. A manager provides intelligent, dynamic routing.