In CrewAI, what do an Agent's role, goal and backstory actually change at runtime?

## The three strings A CrewAI `Agent` requires `role`, `goal` and `backstory`. They are plain strings, and at run time they are interpolated into CrewAI's prompt templates to produce the system prompt for that agent. That is the whole mechanism. There is no hidden registry that reads roles, no scheduler that matches a goal to work, no policy engine that reads a backstory. ## What each one is for in practice **role** is a short identity phrase — "Senior Research Analyst", "Report Formatter". Two things make it more than decoration. First, it is the top line of the persona the model sees, and models respond strongly to a concrete professional identity. Second, it is the agent's address: when one agent delegates, it names the coworker by role string, and crew output and verbose logs identify agents by role. That makes role the one field with a functional side effect, and it is why near-identical roles in one crew ("Researcher" and "Research Analyst") cause real problems — a delegating agent has to pick between them from the role text alone. **goal** is the objective sentence. It is not consumed by any scheduler; it is prompt text that keeps the agent oriented across a long tool loop, when the original task description has scrolled far up the message history. A vague goal ("help the user") gives the model nothing to steer by; a concrete one ("produce a five-bullet summary with a source URL per bullet") measurably shapes intermediate steps, not just the final answer. **backstory** is the largest field and usually the most load-bearing. It is where experienced users put behavioural rules that read as persona but function as constraints: preferred output conventions, tone, what to do when a tool returns nothing, when to stop searching, what never to invent. Because it is unstructured, it is also where prompts rot: a backstory that has grown to a page of contradictory instructions is a common cause of an agent that ignores half of them. ## What they do not do This is the part interviewers actually probe, because it separates people who have shipped a crew from people who have read the landing page. - They do **not** restrict tools. An agent whose backstory says "you only read files, you never write" will happily call a write tool if that tool is in its `tools` list. The allowlist is the list, not the prose. - They do **not** assign work. A task runs on the agent named in the `Task`'s `agent` field (or, under a manager-driven process, whichever agent the manager picks); a matching-sounding goal does not attract work on its own. - They do **not** enforce output shape. Structured output is a task-level contract, not something a backstory guarantees. - They do **not** isolate context. Writing "you are independent" does not stop shared crew state or delegated context from reaching the agent. ## Interpolation All three fields support `{placeholder}` substitution. Placeholders are filled from the `inputs` dictionary handed to the crew at kickoff, which is what lets one agent definition serve many runs — a role of `{topic} Senior Data Researcher` becomes a topic-specific identity per run. The failure mode is a placeholder with no matching input key, which surfaces at kickoff rather than at construction. ## Templates underneath CrewAI wraps these strings in its own scaffolding — the instructions about producing a final answer, the tool descriptions, the formatting rules. An `Agent` can override that scaffolding through `system_template`, `prompt_template` and `response_template` when you need control over the whole prompt rather than the persona slot inside it. Reaching for those is a strong signal you have exhausted what persona text can do. ## Writing them well Keep `role` short, unique within the crew and human-recognisable. Make `goal` a single measurable objective with the output shape implied. Use `backstory` for rules of engagement rather than biography — the model does not benefit from where the persona went to school, but it does benefit from "if the search returns fewer than three credible sources, say so instead of filling the gap". And when behaviour must be guaranteed rather than encouraged, move it out of prose: remove the tool, set the caps, or validate the task's output. ## The honest summary Role prompts are real engineering, but only within their remit. They shape a probabilistic system's tendencies; they never constitute an enforcement boundary. A candidate who says both halves of that sentence is answering the question the interviewer is actually asking.