CrewAI: English Vocabulary for Multi-Agent AI Systems

CrewAI is a Python framework for building systems in which multiple AI agents collaborate to complete complex tasks — each agent playing a specialised role, much like members of a human team. As multi-agent architectures become more common in production AI systems, the ability to describe, design, and debug them in fluent English is increasingly valuable. This guide covers the core CrewAI vocabulary you need to read the documentation, contribute to architecture decisions, and hold precise conversations with colleagues building agentic systems.

Key Vocabulary

Agent — An autonomous unit in CrewAI that is assigned a specific role, goal, and backstory. An agent uses a language model to reason, selects tools to gather information, and produces output that feeds into the broader workflow. Agents are designed to be specialists — a single crew might include a “Research Agent,” a “Data Analyst Agent,” and a “Report Writer Agent.” “We defined a dedicated Compliance Agent with access to the regulatory database tool, rather than giving every agent access to it.”

Task — A discrete unit of work assigned to a specific agent, or delegated by the crew’s manager. Each task has a description (what to do), an expected output (what the result should look like), and optionally a set of tools and a context from upstream tasks. Tasks are the building blocks of a crew’s workflow. “The summarisation task receives the research task’s output as context, so the agent doesn’t need to repeat the information-gathering step.”

Crew — The top-level object in CrewAI that groups agents and tasks together and orchestrates their execution. A crew defines which process to use (sequential or hierarchical), which agents participate, and in what order tasks run. When you call crew.kickoff(), the crew takes over and manages the entire workflow. “We built a separate crew for each report type so that the agent configurations and tools stay focused and easy to maintain.”

Sequential Process — An execution mode in which tasks run one after another, in the order they are defined. The output of each task is passed as context to the next. This is the simplest process type and is appropriate when tasks have clear dependencies and must be completed in a fixed order. “We use a sequential process here because the analysis task can only begin once the data-gathering task has finished.”

Hierarchical Process — An execution mode in which a manager agent coordinates the other agents and decides which tasks to delegate and in what order. The manager can reassign tasks, ask agents to redo work, or synthesise outputs from multiple agents. This mode is more powerful but also more complex to debug. “After switching to a hierarchical process, the manager agent started catching quality issues in intermediate outputs and requesting revisions automatically.”

Manager Agent — A special agent in a hierarchical crew that acts as the orchestrator. It does not necessarily perform domain tasks itself — its role is to decompose goals, assign tasks to the most appropriate specialist agents, and evaluate whether their outputs meet the required standard. “We gave the manager agent a strict quality rubric in its backstory so it knows when to accept an output and when to send it back for revision.”

Memory — CrewAI agents can be equipped with different types of memory to improve consistency across long workflows. Short-term memory stores information within a single crew run. Long-term memory persists information across multiple runs, stored in a local database. Entity memory tracks specific named entities (people, organisations, concepts) mentioned during the workflow. “Enabling long-term memory allowed the research agent to recall findings from last week’s run and avoid duplicating work.”

Kickoff — The method call that starts a crew’s execution. When you call crew.kickoff(), CrewAI begins processing tasks according to the defined process. You can also call crew.kickoff_for_each() to run the same crew against a list of inputs in parallel or in sequence. “We wrapped the kickoff call in a try-except block so that a failure in one pipeline run doesn’t prevent the remaining batch items from being processed.”

Callback — A function you register on a task or agent that is invoked at specific points during execution — for example, when a task completes, when an agent takes a step, or when a tool is called. Callbacks are used for logging, monitoring, and injecting custom behaviour without modifying the core agent logic. “We added a task callback to stream each agent’s intermediate output to our logging service so the team can inspect the reasoning chain in real time.”

Useful Phrases

Engineers building CrewAI systems in English-speaking teams use language like this in design discussions, code reviews, and retrospectives:

• “Let’s delegate the web search to a dedicated agent rather than giving every agent a browser tool — it’s cleaner and easier to rate-limit.”
• “The crew is hallucinating the output structure — we need a stricter expected output definition on that task to constrain the format.”
• “I’d like to add an output parser to that task so we can validate the JSON schema before it’s passed downstream.”
• “We should switch to a hierarchical process here — the sequential process can’t handle the case where one subtask turns out to be unnecessary.”
• “The agent is looping — it keeps calling the same tool repeatedly. Let’s cap max_iter and add a fallback instruction to the task description.”
• “Can you check the entity memory for that run? I think the agent lost track of the company name halfway through the workflow.”

Common Mistakes

Saying “the AI” instead of naming the agent

Non-native speakers often refer to all agents generically as “the AI” in conversations and written summaries — for example, “the AI searched the web and then the AI wrote the report.” In a multi-agent context this creates ambiguity. Always name the specific agent: “the Research Agent searched the web and then the Writer Agent drafted the report.” This habit also makes debugging much easier, because log messages and callbacks are tied to specific agent names.

Confusing “task” and “tool”

In CrewAI, a task is a goal assigned to an agent, while a tool is a capability the agent can invoke (such as a web search, a file reader, or a database query). Non-native speakers sometimes say “I’ll give the agent a summarisation task” when they mean “I’ll give the agent a summarisation tool.” The distinction matters: tasks are defined at the crew level and represent work to be completed; tools are defined at the agent level and represent means of gathering information or taking action.

Misusing “sequential” to mean “slow”

Engineers sometimes use “sequential” as a synonym for “inefficient” when arguing for parallelism. In CrewAI, “sequential process” is a precise architectural term — it means tasks depend on each other’s outputs and must run in order. The opposite is not “fast” or “parallel” but “hierarchical.” When recommending a design change, say “we should consider a hierarchical process to allow the manager agent to run independent tasks concurrently” rather than “we should make it less sequential.”

CrewAI’s vocabulary maps closely onto concepts from human team management — roles, delegation, memory, and feedback loops — which can make it surprisingly intuitive once you know the terms. The more fluently you can describe agent roles, process types, and task dependencies in English, the more effectively you can contribute to design reviews, write clear documentation, and collaborate with the growing community of engineers building multi-agent AI systems.