CrewAI and Ollama: Kubernetes setup and observability

This tutorial builds a BMAD software-delivery crew with eight specialists, then deploys it to Kubernetes. The presenter explains CrewAI's role-based approach alongside LangGraph's graph model and AutoGen's conversational approach. The example uses the open source engine rather than CrewAI's paid hosted platform.

Agents have a role, goal and backstory, with tasks and handoffs defined in YAML. The demo runs tasks sequentially; a hierarchical process lets a manager delegate dynamically. The supplied stack specifies CrewAI 1.15.1, Python 3.10–3.12 and Ollama with qwen3.6. All agents share a local LLM served from another VM, which the cluster must reach over the network.

FastAPI exposes the crew as a service, while a self-hosted CopilotKit interface submits work and receives progress through server-sent events. The presenter reports runs of about 20 minutes in this setup. The entire crew runs in one pod. Persistent storage matters when memory is enabled, and this example's human approval checkpoints work only through the CLI.

The final section addresses telemetry. The presenter finds the default instrumentation insufficient for per-agent token analysis and routes OpenLIT telemetry through an OpenTelemetry Collector to Dynatrace. OpenLIT 1.42.1 leaves the agent-name attribute empty in the tested setup, so the dashboards use span names to identify agents.