Agentic Work

Agentic AI is a first-class layer in FlowLakes. Agents operate the same platform a person does, through one governed control plane. Each concept page below is one piece of that layer.

• How it fits together: the system view — what decides, what explains, what executes, and the one write path they share.
• The agentic loop: trigger, run, plan, approve, and execute, through the one control plane.
• Agents: the configured AI identity, its service account, charter, tools, and autonomy mode.
• Skills and knowledge: versioned prompt fragments and the knowledge layer that assembles a prompt.
• Judgments: single-shot AI decisions with typed contracts and a review queue. The admin page and its tools were retired in 1.29.0; the mechanism is still described here.
• Work items: free-text instructions handled by operator agents. The admin page and its tools were retired in 1.29.0; the mechanism is still described here.
• Operational memory: a per-tenant index of past resolutions that agents recall by keyword so a problem solved once starts the next from precedent.
• Plans and governance: an agent’s approvable proposal and how it is governed.
• Record sessions: the session on every record of a session-enabled table, what wakes it, how it asks, and how a person takes a record over and hands it back.
• Connections: the model this tenant uses, and bounded outbound reach across web, email, Graph, and REST.
• Triggers and events: lifecycle events, observers, and agent handoff.
• Safety model: least privilege, human review, and fail-closed authorization.

New to agents? Start with How it fits together, then the agentic loop and Agents.