Core concepts

The building blocks

• Agent — an LLM-backed actor with a system prompt, a model, a set of actions/skills, and long-term memory. It runs a plan→act→observe loop and calls its own tools.
• Action / Skill — a capability the agent can invoke. An Action is code-backed (a skill with intent); a Skill is a markdown playbook. Both register into one pool.
• Workflow — a graph of nodes (triggers, agents, actions, conditions, variables, handoffs) the engine runs in order, piping each node's output downstream.
• Memory — a workspace-scoped store every agent reads/writes by default (core, no plugin).
• Plugin — a packaged module that registers actions, channels, providers, or UI; activating one adds its actions to the pool.