CrewAI Production Playbook

Look forUse a Flow with start, listen, and router steps so code chooses the next crew, and keep each crew's task small. Builders report agents that collaborate awkwardly without a spec, and a delegated coworker that ran a tool the orchestrator never had.

Look forAdd @persist to keep Flow state in SQLite or your own backend, and turn on checkpointing for long runs. Restore skips completed tasks and reruns the rest. Diagrid's CTO notes that persisted state does not resume by itself, and one issue shows a failed task skipped after restore, so test restore on a forced failure.

Look forGive every tool with side effects an idempotency key. Builders report tools rerun on task retry, failed tool calls retried inside each outer attempt, and a model that reported a search it never ran.

Look forCompare crew.usage_metrics with the provider's billing for a week. Builders found counts doubled when agents shared one LLM instance and Bedrock counts that left out cache reads. Scope each task's context, since default context passing makes a five-agent crew send about six times the tokens of isolated tasks.

Look forSet each agent's model as provider/model-id and supply each provider's key. memory=True embeds with OpenAI text-embedding-3-large by default, so configure another embedder if memory content must stay off OpenAI. An agent with no model set falls back to gpt-4.1-mini in the library's code. Providers without a native integration run through LiteLLM; pin it, since CrewAI pinned LiteLLM 1.82.6 during the March 2026 compromise.

Look forAnonymous telemetry is on by default; set CREWAI_DISABLE_TELEMETRY or OTEL_SDK_DISABLED to stop it, and keep share_crew off. AMP tracing needs CREWAI_TRACING_ENABLED=false. Builders reported telemetry still sent after disabling it and a tracing prompt that added about 20 seconds per run, so confirm with a network capture.

Look forAMP Basic stops at 50 executions a month. Enterprise adds SSO with Microsoft Entra or Okta, role-based access control, PII redaction for traces, and deployment in your VPC, with execution overage quoted per customer. Self-hosting puts workers, storage, and scaling on your team.

Look forPin crewai and its extras. CrewAI shipped five releases between October 14 and October 27, 2026, and builders report upgrades that called tools repeatedly or sent Anthropic-only fields to other providers. Async guardrail retries and concurrent flow setup have open crash reports. Every release requires chromadb 1.5.9, which GitHub advisory GHSA-f4j7-r4q5-qw2c rates critical with no patched version, so clear it with your security team or isolate the service; Headroom and ai-agent-assembly dropped CrewAI from their lockfiles over it.