crewaiinc/crewai

Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.

Structural Verdict

How Data Flows Through the System

Users define agents with roles and tools, create tasks with descriptions and expected outputs, organize them into crews for execution, with optional A2A delegation to remote agents

1. Agent Creation — Define agents with role, backstory, tools, and LLM configuration (config: oauth2_provider, oauth2_client_id, oauth2_domain)
2. Task Definition — Create tasks with descriptions, expected outputs, and assign to agents
3. Crew Assembly — Group agents and tasks into a crew with specified execution process
4. Execution Engine — Crew orchestrator executes tasks using selected process (sequential, hierarchical)
5. A2A Communication — Agents can delegate to remote agents via A2A protocol with authentication (config: provider, client_id, domain +1)
6. Result Collection — Task outputs are collected into CrewOutput with usage metrics

System Behavior

How the system actually operates at runtime — where data accumulates, what loops, what waits, and what controls what.

Data Pools

Feedback Loops

• OAuth2 Token Refresh (retry, balancing) — Trigger: Token expiration or 401 response. Action: Refresh access token using refresh token. Exit: Valid token obtained or max retries exceeded.
• A2A Task Polling (polling, balancing) — Trigger: Task submission to remote agent. Action: Poll task status endpoint for completion. Exit: Task reaches terminal state or timeout.
• Agent Conversation (recursive, reinforcing) — Trigger: Multi-turn A2A conversation. Action: Agent sends message, receives response, may send follow-up. Exit: Max turns reached or agent completes task.

Delays & Async Processing

• A2A Task Execution (async-processing) — Agent waits for remote agent to complete delegated task
• OAuth2 Token Acquisition (async-processing, ~30 seconds timeout) — Authentication setup before first A2A request
• HTTP Request Timeout (rate-limit, ~configurable per client) — Prevents indefinite blocking on network requests

Control Points

• CREWAI_TRACING_ENABLED (env-var) — Controls: Enable/disable telemetry and tracing
• max_turns (threshold) — Controls: Maximum conversation turns in A2A delegation
• oauth2_provider (runtime-toggle) — Controls: Which OAuth2 provider to use for authentication
• DEFAULT_TIMEOUT (threshold) — Controls: API request timeout duration. Default: 600

Technology Stack

Key Components

• Agent (class) — Core agent abstraction with role, backstory, tools, and LLM configuration lib/crewai/src/crewai/agent/core.py
• Crew (class) — Orchestrates multiple agents to execute a sequence of tasks collaboratively lib/crewai/src/crewai/crew.py
• Task (class) — Individual work unit with description, expected output, and assigned agent lib/crewai/src/crewai/task.py
• Flow (class) — Higher-level workflow abstraction for complex multi-step processes lib/crewai/src/crewai/flow/flow.py
• A2AConfig (config) — Configuration for Agent-to-Agent protocol communication with authentication lib/crewai/src/crewai/a2a/config.py
• LLM (class) — Language model abstraction with provider-agnostic interface lib/crewai/src/crewai/llm.py
• DeployCommand (cli-command) — Handles deployment of crews to CrewAI Cloud platform lib/crewai/src/crewai/cli/deploy/main.py
• Oauth2Settings (config) — OAuth2 authentication configuration for enterprise/cloud features lib/crewai/src/crewai/cli/authentication/main.py
• ExtensionRegistry (utility) — Registry for A2A protocol extensions that customize agent behavior lib/crewai/src/crewai/a2a/extensions/base.py
• Knowledge (class) — Knowledge base abstraction for providing context to agents lib/crewai/src/crewai/knowledge/knowledge.py

Sub-Modules

Configuration

lib/crewai-tools/src/crewai_tools/rag/base_loader.py (python-pydantic)

• content (str, unknown) — default: Field(description="The text content of the source")
• source (str, unknown) — default: Field(description="The source of the content", default="unknown")
• metadata (dict[str, Any], unknown) — default: Field(
• doc_id (str, unknown) — default: Field(description="The id of the document")

lib/crewai-tools/src/crewai_tools/rag/core.py (python-pydantic)

• id (str, unknown) — default: Field(default_factory=lambda: str(uuid4()))
• content (str, unknown)
• metadata (dict[str, Any], unknown) — default: Field(default_factory=dict)
• data_type (DataType, unknown) — default: DataType.TEXT

lib/crewai-tools/src/crewai_tools/tools/arxiv_paper_tool/arxiv_paper_tool.py (python-pydantic)

• search_query (str, unknown) — default: Field(
• max_results (int, unknown) — default: Field(

lib/crewai-tools/src/crewai_tools/tools/brightdata_tool/brightdata_dataset.py (python-pydantic)

• API_URL (str, unknown) — default: "https://api.brightdata.com"
• DEFAULT_TIMEOUT (int, unknown) — default: 600
• DEFAULT_POLLING_INTERVAL (int, unknown) — default: 1

Related Data Pipeline Repositories

Frequently Asked Questions

Analyzed on March 31, 2026 by CodeSea. Written by Karolina Sarna.