LANDMARK · 8 MIN · BUILD

Agent loops

After this landmark, you can describe the plan–act–observe loop an agent runs, and name where that loop needs a stop condition or human checkpoint.

This concept is shared, but the Everyday lens is less central here.

You can still explore it. We’re showing the shared explanation and a related practical view without hiding the knowledge.

SHARED FOUNDATION

An agent’s core mechanism is a loop: the model decides an action, a tool executes it, the result comes back as new context, and the model decides the next action; repeated until the task is done or the loop is stopped. That loop is powerful because it lets a model handle multi-step tasks without a human scripting every step, and risky for the same reason: a wrong turn early in the loop can compound, since each step’s context includes the (possibly wrong) results of the last.

Designing the loop means deciding its stop conditions, its maximum steps, and where a human reviews before an irreversible action goes out.

This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.

THE LOOP THIS RING KEEPS COMING BACK TO
Every pass around the loop feeds the last step's result back in as new context, which is why a wrong turn early on compounds, and why the checkpoint sits before Act rather than after it.
EVERYDAY LENS

What this looks like for you

When an ‘agentic’ assistant is working through a multi-step task for you, notice whether it shows its steps as it goes or only the final result; visibility into the loop is what lets you catch it going sideways.

MAKE A DECISION

An agent is booking a multi-city trip: search flights, pick options, book, confirm with the user. Where’s the checkpoint that matters most?

CARRY THISFor one agent or automated workflow you use, find its irreversible step (send, pay, delete, publish). Is there a checkpoint before it, or does the loop run straight through?
CONCEPTS IN THE INDEX