How to Deploy CrewAI to Production

When you shift CrewAI agent crews from notebooks to production, their behavior changes. They need memory that lasts across runs, storage for outputs, and team-shared access. Without these, state resets after each task, limiting crews to one-offs. Production agents tackle ongoing jobs, such as customer support chats or data pipelines. Fast.io workspaces support this: agents store files, search via RAG, and start tasks with webhooks. Teams can scale CrewAI this way. Check Fast.io Workspaces, Fast.io Collaboration, and Fast.io AI. Example: A research crew handles daily market data, saves to Fast.io. Agents require API keys for MCP.

Follow this step-by-step checklist for CrewAI production. 1. Install with pip install 'crewai[tools]'. Pin versions like crewai==1.9.3 in requirements.txt. Use UV for CI/CD speed.

Services: Redis for memory, Postgres for history, LLM like OpenAI/Anthropic, Serper for search. Try Upstash or Supabase managed options. 3.

Dockerfile: dockerfile FROM python:3.11-slim WORKDIR /app COPY requirements.txt. RUN pip install -r requirements.txt COPY.. CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000"] Add health checks. 4. docker-compose.yml: yaml services: app: build:. ports: ["8000:8000"] depends_on: [redis, postgres] redis: image: redis:7-alpine postgres: image: postgres:16 environment: POSTGRES_DB: crewai 5.

Env Vars: LLM keys, DB URLs, memory settings. 6.

Observability: LangSmith callbacks. 7.

Local Test: docker compose up. 8.

Deploy: Render, Railway, Fly.io. 9. 10. CI/CD: GitHub Actions with ruff lint, pytest, bandit scans. 11.

Health: /health endpoint for agent/DB status.

Single-container prototypes work fine, but production needs auto-scaling, managed DBs, and reliability. Compare options:

Railway suits most CrewAI apps with built-in scaling, Postgres, Redis. Add via railway add.

For Kubernetes, use this Deployment + HPA:

apiVersion: apps/v1
kind: Deployment
spec:
  replicas: 3
  template:
    spec:
      containers:
      - name: crewai
        resources:
          requests:
            cpu: "100m"
            memory: "256Mi"
        envFrom:
          - secretRef:
              name: crewai-secrets
---
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
spec:
  scaleTargetRef:
    kind: Deployment
    name: crewai
  metrics:
  - type: Resource
    resource:
      name: cpu
      target:
        type: Utilization
        averageUtilization: 70

Serverless: Lambda for <15min tasks, SQS queues, DynamoDB state.

Celery + Redis for task queues: celery -A tasks worker -l info -c 4.

LLM handling: tenacity backoff, circuit breakers.

Fast.io takes file I/O load off agents via MCP uploads.

Deploy CrewAI Agents Today

50GB free storage, 251 MCP tools, RAG. No credit card. Works for CrewAI production. Built for crewai production deployment workflows.

Sign Up Free

CrewAI lacks native file persistence. Fast.io workspaces handle it. Agents create orgs/workspaces via API and upload via MCP tools over HTTP/SSE, no local storage needed. Intelligence Mode indexes files for RAG. Agents ask "Summarize last report" and get cited answers. Free plan: 50GB storage, large files, 5000 credits/month, no card. Ownership transfer passes control to humans, agents keep admin. Example: Agent sets up client data room, transfers to PM. Human accepts by email.

Fast.io logs track uploads, views, changes. Webhooks alert on events and trigger crews. Workflow: CSV upload → webhook → research crew → report back. Pair with LangSmith for traces. 01/task. File locks stop multi-agent conflicts (10min timeout).

Memory loss? Check Redis/Postgres connections and health. LLM limits? Tenacity retries + Celery queues. File conflicts? Lock via Fast.io before writes. Scaling hangs? DB pool exhaustion, add pgbouncer. Start with logs: crewai --verbose, dashboards, LangSmith. 1%.