Install
pip install "crewai[tools]"Role-based autonomous agent teams with Flows, MCP, and checkpoint resume.
Role-based multi-agent orchestration. Define agents with roles, goals, and backstories; chain them into crews with sequential or hierarchical processes. CrewAI Flows add event-driven workflows; checkpoints let you resume long-running crews.
Install
pip install "crewai[tools]"Version
v1.15.41 · checked 2026-11-19 · Python 3.10–3.13
Best for
Team simulations, research pipelines, content creation, analyst crews.
These were found by tracing what agents and tasks actually receive, and each is covered where it applies.
Agent(guardrail=...) only applies to Agent.kickoff(). In a crew, use Task.guardrail (Agent API).Task(response_model=...) leaves TaskOutput.pydantic empty; use output_pydantic for a parsed result (Task API).{variables} (Task API).{placeholders} are not interpolated into a manager_agent’s own text (Multi-agent patterns).step_callback sees tool steps only on the text (ReAct) path; with native function calling it fires once with the final answer (Agent API).knowledge/, and each task with knowledge makes an extra LLM call (Knowledge).Task(input_files=...) needs crewai[file-processing] and File(...) objects (Task API).Memory.forget(scope=..., older_than=...) deletes newer records (seen in 1.15.28, still present in 1.15.41), and forget(record_ids=[]) deletes everything (Memory).crewai train asks you for feedback on every task (CLI).crewai create crew now runs an interactive wizard and writes crew.jsonc + agents/*.jsonc. Pass --classic for the Python/YAML @CrewBase layout this guide uses. Tools, skills and templates are scaffolded through the same command (crewai create tool|skill|template); the older crewai tool create / crewai template add forms are deprecated.crewai create flow --declarative, crewai run --definition <file>; conversational flows are stable since 1.15.18. crewai run runs flows as well as crews.crewai eval — grades the last traced run through CrewAI AMP and exits non-zero on a failed gate; --models compares models on a deployment.AgentExecutor is now the default executor_class; new skills= and tool_failure_policy= fields; tool failures are surfaced rather than reported as success.AgentExecutor moved to crewai.agents.agent_executor and CrewAgentExecutor was removed; new Agent.use_better_compaction (large native tool results are written to a temporary file the agent reads in bounded sections); a run now stops with ToolExecutionFailedError on the fifth identical tool call; crewai eval runs the crew first when there is no traced run, and gains crewai eval export; crewai-tools adds AzureDocumentDBVectorSearchTool. The two Memory.forget bugs above are not fixed.crewai-cli distribution (installed with crewai; crewai.cli is a deprecated alias for crewai_cli). Provider SDKs and LiteLLM are optional extras (crewai[anthropic], crewai[google-genai], crewai[bedrock], crewai[litellm]).gpt-5.4-mini and embedder text-embedding-3-large; stores built with the old 1536-dimension default need a reset or an explicit embedder (see Memory).from_checkpoint on Crew.kickoff / Flow.kickoff / Agent.kickoff — resume with from_checkpoint=CheckpointConfig(restore_from=...).crewai checkpoint list|info|resume|diff|prune.reasoning_tokens, cache_creation_tokens on CrewOutput.token_usage.from crewai import Agent, Task, Crew, Process, LLM
llm = LLM(model="openai/gpt-6-luna")
analyst = Agent( role="Research Analyst", goal="Analyse topics thoroughly", backstory="Expert analyst with 10 years of experience", llm=llm,)
task = Task( description="Research AI advancements in 2026", expected_output="Concise analysis of key developments", agent=analyst,)
crew = Crew(agents=[analyst], tasks=[task], process=Process.sequential)result = crew.kickoff()print(result)Ready? Start Zero → Hero Step 1 →
| Date | Version | Changes | Reviewer |
|---|---|---|---|
| 2026-11-11 | 1.15.39 | Bumped from 1.15.30. The built-in default model changed from gpt-4.1-mini to gpt-6-luna in 1.15.38 (checked crewai.constants.DEFAULT_LLM_MODEL and the planner default); updated the four pages that state it. Added crewai eval --no-open. No breaking changes in 1.15.31–1.15.39 changelog. | Claude |
| 2026-10-25 | 1.15.30 | Bumped from 1.15.28. Re-ran the version-specific claims (default model, tool caching matrix, usage-cap quirk, response_model, task-event agent_role, deploy validate on declarative flows, flow.plot location) and the offline pytest examples: all unchanged. Both Memory.forget bugs re-tested and still present. Documented the AgentExecutor move, use_better_compaction, the repeated-call stop and crewai eval export. | Claude |
| 2026-10-25 | 1.15.28 | Deep-dive audit: re-ran every page’s examples, traced what agents and tasks receive, and corrected the Agent, Task, Memory, Knowledge, Tools, LLM, CLI, Production, Multi-agent, Recipes and Diagrams pages (agent-level guardrails, response_model, task context, hierarchical manager behaviour, tool caching and failure policy, step_callback, knowledge paths and input_files, memory in crews, forget, training, replay). Added a local MCP example, a ToolFailure section, an offline Memory example and PlanningConfig fields. | Claude |
| 2026-10-21 | 1.15.28 | Rewrote the Flows, Production, Recipes, Diagrams, Multi-agent and Comprehensive pages against 1.15.28 (structured flow state, routers that pass 1.15’s definition checks, and_/or_, @persist, @human_feedback, conversational and declarative flows; event listeners, OpenTelemetry, hooks, checkpoints, offline tests with a stub LLM, crewai deploy validate). Fixed reference-page errors (embedder model_name, aadd on custom knowledge sources, MCP filter signature, guardrail return values, default task context, CrewOutput.to_dict, Qdrant Edge extra). Links made root-absolute. | Claude |
| 2026-10-21 | 1.15.28 | Re-verified the whole guide against 1.15.28 in a clean venv (imports, constructor keywords, CLI --help). Current-version mentions now come from the shared versions file. Documented JSON-first scaffolds, crewai eval, crewai_cli split, new Agent fields and defaults, Memory default embedder change; fixed memory_tools_for → create_memory_tools, A2AClientConfig(endpoint=...), crewai_tools imports; removed the fabricated crewai.amp examples. | Claude |
| 2026-05-29 | 1.14.6 | Version bumped 1.14.5 → 1.14.6; “What’s new” heading updated to v1.14.6; v1.14.6 bullet added; comprehensive guide header updated; revision history entry added. Agent, Task, Crew, Process, Flow verified against installed crewai==1.14.6 (.routine-envs/check-0529-crewai); no DeprecationWarning emissions. | Claude routine |
| 2026-05-19 | 1.14.5 | Version bumped 1.14.4 → 1.14.5; “What’s new” section updated; comprehensive guide header updated; revision history entries added. Agent, Task, Crew, Process, Flow verified against installed crewai==1.14.5 (.routine-envs/check-0519-crewai); no DeprecationWarning emissions. | Claude routine |
| 2026-05-01 | 1.14.4 | Version bumped 1.14.3 → 1.14.4; “What’s new” section updated; revision history entry added. Agent, Task, Crew imports verified against installed 1.14.4 (.routine-envs/check-crewai-0501). | Claude routine |
| 2026-04-25 | 1.14.3 | Pre-release 1.14.3a2 confirmed stable as v1.14.3; version card updated from pre-release to stable; revision history entry added. Symbols verified against installed crewai 1.14.3 (.routine-envs/crewai-py-0425). | Claude routine |
| 2026-04-22 | 1.14.3a2 | Added 7 source-verified reference pages: Agent API, Task API, Memory, Knowledge, Tools, LLM providers, CLI. Corrected checkpoint import path (crewai.state.provider.sqlite_provider), field name (checkpoint=, not checkpoint_config=), and resume API (from_checkpoint=CheckpointConfig(restore_from=...)). Added errata banner to the AMP Suite section (imports fabricated; use CLI). | Claude (Opus 4.7) |
| 2026-04-21 | 1.14.2 | Index redesigned with Zero → Hero + Jump-to-topic grid. | — |
| 2026-04-17 | 1.14.2 | Checkpoint forking; from_checkpoint; checkpoint CLI; template CLI; reasoning_tokens. | — |
| 2026-04-07 | 1.14.0 (yanked) | Checkpoint system; Pydantic structured outputs; SSRF protections; RBAC; MCP custom servers. Re-released as 1.14.1 on 2026-04-08. | — |
| 2025-11 | 1.5.0 | CrewAI Flows; MCP integration; memory systems; hierarchical processes. | — |