crewAI 1.15.2 rolls out dynamic LLM models and smarter flow skills
55K-star framework now lets agents pull latest models on the fly and handle feedback in real time.
crewAI v1.15.2 introduces several powerful features for building agentic workflows. The headline addition is dynamic LLM model pulling in the crew wizard, allowing agents to automatically load the latest models without manual configuration. Inline skill definitions let developers embed custom capabilities directly into flow definitions, while templated Flow action inputs make it easier to parameterize agent behavior. A new text helper for CEL prompts and a generated Flow Definition authoring skill further reduce boilerplate for complex patterns.
On the infrastructure side, the update adds a stream frame protocol for flows, enabling real-time data streaming between agents. The AgentExecutor now includes message setup and feedback handling, giving agents a structured way to receive and act on user or system feedback. Repository agents can be added to flow definitions, and the framework types tools and apps via CrewDefinition. Bug fixes focus on caching stability (keying model-catalog cache by exact API key with shorter TTL), resolving pip-audit vulnerabilities, and ensuring version consistency for flows. With 7.8K forks and a vibrant community, crewAI continues to push the boundaries of multi-agent coordination.
- Dynamic LLM model pulling in the crew wizard eliminates manual model updates for agents.
- Inline skill definitions and templated Flow action inputs simplify custom agent behavior authoring.
- AgentExecutor now supports message setup and feedback handling, enabling real-time interactive agents.
Why It Matters
Empowers developers to build adaptive, real-time multi-agent systems with less boilerplate and smarter model integration.