crewai

CrewAI enables defining autonomous agents with specific roles, goals, and backstories that collaborate on complex tasks through a Crew abstraction. Each agent is instantiated with an LLM provider, tools, and memory context, then orchestrated via a task queue where the framework automatically routes work based on agent capabilities and task dependencies. The Crew class manages agent lifecycle, handles inter-agent communication, and enforces sequential or parallel task execution patterns with built-in retry logic and error recovery.

CrewAI Flows provide a decorator-based, event-driven orchestration layer built on top of Crews, enabling complex workflows where steps are triggered by state changes rather than explicit sequencing. Flows use Python decorators (@flow, @listen_to) to define state machines where each decorated method represents a workflow step that can emit events, listen for upstream events, and compose Crews as sub-workflows. The framework manages state persistence, event routing, and visualization of the entire workflow DAG.

CrewAI provides the crewai-files package for agents to read, write, and process files and documents. The package includes tools for file operations (read, write, delete, list), document parsing (PDF, DOCX, TXT, JSON), and file-based memory operations. Files are managed in an agent-scoped workspace, enabling agents to work with documents without direct filesystem access. The system integrates with the memory architecture to enable semantic search over document contents.

CrewAI AMP is the enterprise deployment platform providing managed hosting, control plane, monitoring, and governance for deployed crews. AMP handles agent lifecycle management, automatic scaling, environment variable injection, secret management, and integration with enterprise identity systems (SSO). The platform provides a web UI (Crew Studio) for managing deployed agents, viewing execution logs, and triggering manual runs. AMP integrates with CrewAI's marketplace for discovering and deploying pre-built agents.

CrewAI provides an evaluation framework for testing agent behavior, measuring performance against benchmarks, and comparing agent configurations. The framework enables defining test cases with expected outputs, running agents against test suites, and collecting metrics (accuracy, latency, cost). Evaluation results can be compared across agent versions, LLM models, or tool configurations, enabling data-driven optimization. The framework integrates with the observability system to capture detailed execution traces for failed tests.

CrewAI provides an agent skills system enabling agents to be composed from modular, reusable skill components. Skills are Python classes that encapsulate specific capabilities (e.g., 'web research', 'code analysis', 'report writing') and can be attached to agents at instantiation. Skills have their own tools, memory, and execution context, enabling complex agent behaviors to be built from simple, composable pieces. Skills can be versioned, shared across agents, and discovered through the marketplace.

CrewAI abstracts over multiple LLM providers (OpenAI, Anthropic, Gemini, Azure, Bedrock, Ollama) through a unified LLM class that normalizes provider-specific APIs into a common interface. The framework handles provider-specific message formatting, function-calling schema translation, and streaming response handling. Each provider implementation extends a base LLM class and implements hooks for pre/post-processing, enabling agents to seamlessly switch providers or use provider-specific features without code changes.

CrewAI provides a tool registry system where tools are defined as Python callables with type hints, automatically converted to provider-specific function-calling schemas (OpenAI, Anthropic, Gemini formats). The framework supports both native Python tools and Model Context Protocol (MCP) tools, normalizing both into a unified tool interface. Tools are bound to agents at instantiation, and the agent's LLM automatically invokes them based on function-calling responses, with built-in error handling and result injection back into the agent's context.

CrewAI implements a unified memory system where agents store experiences, learnings, and context in a multi-tier architecture: short-term memory (current task context), long-term memory (embeddings of past interactions), and entity memory (extracted knowledge about specific topics). The system uses embeddings (via configurable providers like OpenAI, Ollama) to enable semantic search and recall of relevant past experiences. Memory operations (store, recall, consolidate) are scoped to individual agents or crews, with automatic consolidation that summarizes old memories to prevent context bloat.

CrewAI implements an Agent-to-Agent (A2A) protocol enabling agents to directly request work from other agents within a crew, bypassing the task queue. When an agent determines it needs specialized work from another agent, it can invoke the A2A protocol to send a structured request, receive the result, and continue its own reasoning. This enables dynamic task delegation based on agent reasoning rather than static task definitions, with built-in request/response formatting and error handling.

CrewAI provides task-level guardrails that enforce constraints on agent execution, including output validation, structured output schemas, and execution callbacks. Tasks can define expected output formats (via Pydantic models or JSON schemas), and the framework validates agent outputs against these schemas before marking tasks complete. Guardrails include pre-execution hooks (to validate task setup), post-execution hooks (to validate outputs), and custom validation functions that can reject outputs and trigger retries.

CrewAI includes native OpenTelemetry (OTEL) instrumentation that automatically traces agent execution, task completion, LLM calls, and tool invocations. The framework exports traces to OTEL collectors, and integrates with third-party observability platforms (Datadog, New Relic, Langsmith, etc.) via OTEL exporters. Traces include structured metadata (agent name, task description, LLM model, tool names) enabling detailed analysis of agent behavior, latency bottlenecks, and cost attribution.

CrewAI provides a CLI tool that scaffolds new crew, flow, and tool projects with pre-configured project structure, dependencies, and templates. The CLI uses UV for dependency management and includes commands for creating projects, running crews locally, and deploying to CrewAI's managed platform (CrewAI AMP). Deployment commands handle authentication, environment variable injection, and control plane integration, enabling one-command deployment of agent systems.

CrewAI provides LiteAgent, a lightweight agent variant optimized for resource-constrained environments (edge devices, serverless functions, low-memory systems). LiteAgent reduces memory footprint and execution latency by stripping non-essential features (memory consolidation, advanced hooks, complex reasoning) while maintaining core agent capabilities (tool calling, task execution, basic memory). LiteAgent is compatible with standard Crew orchestration, enabling hybrid crews with both full agents and lightweight agents.