How does the agent loop work?

Put the ingredients together and you get the loop that defines an agent: reason, act, observe, repeat.

The model thinks about the goal and picks a next step — that’s the “reason.” It calls a tool: read a file, run a search, send a request. That’s the “act.” The result comes back and the model reads it: the “observe.” Then it decides whether the goal is met, whether the plan needs to change, or whether it needs another pass through the same three steps. The pattern has a name in the research literature: ReAct, short for “reason plus act,” from a 2022 paper showing that interleaving a model’s reasoning with its actions lets each one steer the other — the reasoning shapes what the model does, and what comes back from doing it reshapes the reasoning.^[1]

What ends the loop matters as much as what runs it. A well-built agent stops when it has real evidence the goal is met: a test passes, a record exists, a page returns the right content. Not the moment it asserts that it’s done. OpenAI’s guidance for production agents is explicit about this: the system should recognize when a workflow is actually complete and correct its own actions if something’s off, rather than stop early or keep going past the point of usefulness.^[2] Left unchecked, the loop can also run too long — without a step budget or a check-in point, an agent can spin through many rounds chasing a goal that was ambiguous from the start. That’s what rules and guardrails and runtime guardrails are for.

The fragile part is usually the first step: turning a vague goal into the right sequence of moves. Physea’s Liminality MCP does that part as a tool an agent can call, returning a grounded, reusable route instead of a one-shot guess.