Agents

Building LLM-powered agents on Flux — the agent primitive, its capabilities, the building blocks it composes from, integrations, and proven patterns.

Agents is the section for building LLM-powered reasoning loops on top of Flux. An agent is a specialization of a task: it runs a tool-using loop against a provider, but inherits the workflow’s execution context, event log, and replay-based persistence — so an agent run is durable the same way a workflow run is.

Foundations covers what an agent is, how it sits inside a workflow, and how to pick a provider. Start with The agent primitive.

Capabilities is the feature surface: tools and structured output, memory, dreaming and reflection, sub-agents, reasoning models, streaming, and tool approval. Start with Tools and structured output.

Building blocks covers the pieces an agent composes from — the agent harness, agent skills, agent plans, and system tools. Start with Agent skills.

Integrations covers wiring agents to outside frameworks and protocols: MCP, parallel tool execution, CrewAI, and LangChain. Start with MCP integration.

Patterns is the worked-example section: multi-agent code review, RAG with durable retrieval, human-in-the-loop approval, and long-running research agents. Start with Multi-agent code review.