Skip to content
flowlakes FlowLakes documentation home

The agentic loop

Agentic work in FlowLakes follows one shape, whoever starts it. A run begins from chat, an agent trigger, an action button, or a workflow, and the agent works in a loop: it reads what it needs, and when it wants to change something, that write is gated by the agent’s autonomy mode. Reads run on their own. Writes either execute, wait for approval, or become a proposal for a human, depending on the mode. Every run ends in one of three states (completed, awaiting approval, or failed) and every step is anchored to a record you can audit.

When you’d use it

  • Understand what happens when you dispatch an agent, from trigger to outcome.
  • Predict how a run ends so you know whether to expect a result, a proposal, or a failure.
  • See where the human sits in the loop before you let an agent act.
  • Follow a run’s audit trail back to the record it touched.

The loop, end to end

How the agentic loop runs from trigger to outcomeA trigger starts a run, which resolves knowledge and filters tools before the model-call and tool-dispatch loop. Read tools execute directly, while write tools are gated by autonomy mode, running under autonomous mode or going to approval or a proposal under bounded or propose mode. The run ends completed, awaiting approval, or failed.

read tool

write tool

autonomous

bounded or propose

trigger

run starts

load knowledge, allowed tools

think, then call a tool

run it

autonomy mode

approval or proposal

completed, awaiting approval, or failed

The pieces of the loop each have their own page: the Agent is the configured identity, its Skills and knowledge supply its instructions, a Plan is the approvable proposal, External capabilities reach outside the platform, Triggers and events start and chain runs, and the Safety model governs what may run unattended.

Good to know

  • A run always ends in one of three states: completed, awaiting approval, or failed. There is no silent hang.
  • Reads run on their own; writes are gated. What happens to a write depends entirely on the agent’s autonomy mode.
  • The same loop, permissions, and audit apply whether a person or a trigger started the run.

How it connects

The agentic loop is the frame the other agentic pages fit into. It does not choose an AI provider, own approval routing, or keep chat history as your data. Governance of what may execute lives in the Safety model, and who approves lives in Approvals.