CrewAI 1.15

Apply the Python 3 baseline first. These rules cover CrewAI 1.15 agents, tasks, tools, crews, delegation, flows, state, memory, knowledge, observability, and deployment.

Agents and tasks

• Define agent roles, goals, backstories, models, and task descriptions in trusted application configuration.
• Keep user, retrieved, memory, and tool content separate from authoritative task instructions.
• Give each agent only the tools required for its assigned tasks and do not treat its role text as authorization.
• Set max_iter, max_execution_time, max_retry_limit, and max_rpm explicitly; CrewAI 1.15.21 defaults to 25 iterations, no time or RPM limit, and two retries.
• Validate expected and structured task output independently before persistence or downstream use.
• Constrain Task.output_file and its interpolated inputs to an authorized output directory; task completion can create directories and overwrite the resolved file.

Tools and execution

• Register tools from a fixed trusted registry with narrow input schemas and bounded output.
• Keep the default Crew(cache=False) and set Agent(cache=False) for authorization-dependent, tenant-sensitive, or volatile tools; explicitly passing Agent(cache=True) or a cache handler can enable caching independently of the crew. Agents otherwise default to cache participation, and identical tool arguments can return the first result without re-executing the tool, so never share that cache across trust domains.
• Validate tool arguments and reauthorize the exact operation against current principal, tenant, and resource