Back to the changelog

• Feature

  Flow Agent: start flows in natural language

  A new assistant understands a task written in plain language (*"a new customer Acme just arrived, contact m@acme.it, start the onboarding"*), picks the right flow among the startable ones on its own, collects the missing fields with targeted questions (slot-filling) and proposes the start. The item is created **only after explicit confirmation**, with full audit: the initial transition is tracked with origin `Agent`. *(Commercial edition.)*

• Feature

  Answers on existing items + flow map

  Ask *"how far along is Acme's onboarding?"* and the agent shows a clickable progress card in chat; click it to open the item detail with tasks and transitions. When you ask how a flow is built, it offers a card that opens the read-only flow map (mini-viewer), with an **"Open in designer"** shortcut for Admins / Flow Designers.

• Feature

  Company knowledge (Agent Policy)

  A new **Agent Policy** tab in Settings lets you describe who you are, how you name things and your base policies. The agent uses this context — together with the catalogue of startable flows — to choose well and speak in the company's language.

• Integration

  Three channels, one brain

  The same logic is exposed as a chat in the backoffice, as a REST API (`/agent/*`) and as an MCP tool (`flow_agent`) for external LLM agents — all governed by deterministic guardrails (RBAC, per-flow start permission, confirmation, field validation). The Flow Agent is **off by default** (free behaviour) and enabled via `FLOW_AGENT_ENABLED=true`.