How agents think and act

Agents work in a loop of plan, act and observe. See how each turn of the loop moves a task forward and where it can go wrong.

Scroll, swipe or press ↓ to begin

Every agent runs the same loop: plan, act, observe

Understanding this loop is the fastest way to predict what an agent will do, and why it sometimes gets stuck.

The model reads the goal and everything it knows so far, then picks one next step.

It calls a tool, for example searching the web or updating a row in a sheet.

It reads what the tool returned and adds it to what it knows. That result feeds the next plan.

Each turn of the loop feeds the result of one action into the next plan. The loop repeats until the goal is met or a limit is reached.

Here is the loop for “find this week’s unpaid invoices”:

Plan    → search the invoices sheet for "unpaid"

Act     → sheet.search("unpaid")

Observe → 4 rows found, 1 dated last month

Plan    → keep only this week's rows

Act     → filter(rows, week = this week)

Observe → 3 invoices. Goal met, stop.

Agents fail in predictable ways:

Clear instructions and a small set of well-described tools prevent most of these problems. Set a limit on the number of turns so a stuck agent stops.