How CrewAI's Multi-Agent 'Conversations' Actually Work: Function Calls, Not Dialogue

A developer tested CrewAI, an open-source Python framework, by setting up three AI agents — a Research Analyst, Content Writer, and Editor — to collaboratively produce a factual brief on a home AI server topic. The agents appeared to converse, with the Editor catching a pricing error and sending a paragraph back to the Writer for correction after consulting the Analyst. However, examination of the activity log revealed that inter-agent communication is not dialogue but function calls: CrewAI provides agents with tools like ask_question_to_coworker and delegate_work_to_coworker, which trigger fresh model calls rather than live exchanges. The entire three-agent run completed in 118 seconds, consuming around 18,500 tokens across input and output. Understanding this distinction matters for developers, as the quality of agent interactions depends on how well arguments are structured in those function calls, not on any emergent conversational ability.