A three-agent CrewAI run spent 44.6 of its 118 seconds inside coworker tool calls

When a CrewAI agent is allowed to delegate, the framework registers two extra functions on it, ask_question_to_coworker and delegate_work_to_coworker ^[15]. They arrive the same way a web search or a file reader would ^[15]. Call the first with a coworker's name, a question and some context, and CrewAI starts a fresh model call as that coworker; whatever comes back reaches the caller as the tool's result ^[16]. "A conversation between agents isn't a conversation. It's a function call with a job title," the post's author wrote ^[17]. So the Editor's challenge to the $1,600 figure was a tool invocation, and the Analyst's rebuttal was that tool's return value.

The three logged durations add up to 44.6 seconds ^[21]. That is about 38 percent of the 118-second run ^[22]. Each of those calls is a model invocation on top of the three sequential tasks, so the run made at least six ^[23]. The post reports 14,609 tokens in and 3,928 out for the run as a whole and does not break them down by agent or by call ^[10][27].

The model split is the deliberate part. Analyst and Writer on Qwen 3 235B through OpenRouter, Editor on Claude Opus 4.8 direct from Anthropic ^[3]. "The cheap model does the volume; the expensive one makes the judgement calls," the author wrote ^[4]. Because the coworker's own model generates the answer, the Editor's two questions ran as the Analyst ^[16]. How well that transfers depends on how many rounds your reviewer opens, and this run logged three calls and one caught error ^[14][13].

CREWAI_DMN=true does not appear in crewai --help; the author found it in the CLI's source, where it switches off every interactive prompt ^[6]. It also stops the scaffold writing your .env, so the keys are yours to add ^[7]. The --provider flag only chooses the starting model written into each agent's file ^[8]. The project itself is plain JSON, one file per agent and one crew.jsonc for the tasks ^[9].

The check on the correction happened inside the models. The crew had no web access, so the Analyst confirmed the new numbers from the same kind of training data that produced the $1,600 ^[19]. "A second model agreeing is a second opinion, not a second source," the author wrote ^[20]. The $2,500 to $3,500 range and the claim that a single 24 GB card cannot run a 70-billion-parameter model at usable speed are plausible and unverified ^[12][19].

Nothing is shared between agents except what one call gives the next, and the task handoffs work the same way, each task's output pasted into the next task's prompt ^[18]. The Editor only knew about the $1,600 because it sat in the draft it was handed, and the Analyst could only answer what the question contained ^[29].