CrewAI

CrewAI's tool-usage layer catches the exception and hands the error back to the agent as that tool call's observation, so the kickoff continues and the LLM can retry or change course. The cost is extra model turns, so bound it and return short, actionable error text yourself.

CrewAI's built-in memory is scoped to a storage location, not to a user, so one shared directory means every tenant's content is retrievable by every run. Isolate by giving each tenant its own storage root and process, delegate to a user-scoped external memory provider, or run with memory disabled.

Rarely. Setting human_input=True pauses the task after the agent's answer and blocks on console input for feedback, so it fits an operator running a crew interactively. A server or worker process has no console, so production approval gates belong outside the crew run.

planning=True makes CrewAI call a planning LLM before execution, produce a step-by-step plan for the crew's tasks, and append that plan to each task's description. It costs an extra LLM call per kickoff and can bake in wrong assumptions.

Use MCPServerAdapter from crewai_tools: construct it with the server's connection parameters, and it exposes the server's tools as CrewAI tools you pass into Agent(tools=...). It holds a live connection, so use it as a context manager or call stop() in a finally block.

Give CodeInterpreterTool to one narrowly scoped agent, never to agents that also ingest untrusted text. It executes generated Python inside a Docker container by default; unsafe_mode=True runs it in the host process, which is a development-only setting.