Explain like I'm five · AI

What is an Agent?

AI agents, told as a story about a parrot that, given a goal, plans its own errands, uses the toolbox, and picks the next step from each result.

get my trip next week readyTO DOfind a flightbook a hotelfit the meetingscheck weather, packit takes many stepsthis is not a one-answer job

The guest asks to get a trip ready. This is not a one-answer job.Find a flight, book a hotel, fit the meetings, check the weather and pack. Many steps — and each result changes the next step.

buy the ticket from the airlinesays it once and stops — finds nothing, buys nothingTO DOflight ×hotel ×meetings ×weather, bag ×nothing got donea one-question, one-answer parrot can't do a many-step job

A one-question, one-answer parrot can't do a many-step job.The words sound fine, but nothing is found or bought. Even with a toolbox, using it once and stopping leaves the list untouched.

ERRANDS (its own)✓ find a flight✓ book a hotelnow: fit the meetingsnext: packing listit picks the next step from the resultnext: the calendar tool!toolboxtrainer (runs it)the result comes back, ✓ on the listan agent is a parrot that, given a goal, plans its own errands and uses tools

An agent is a parrot that, given a goal, plans its own errands, uses tools, and picks the next step from each result.Look, think, act, look again. Give it the goal and the parrot picks the steps. The trainer still runs the tools, and a fence caps the steps and the bean budget.

Give the goal, leave the stepsthe parrot writes the list
It needs a toolboxit asks by note
Look and adjust each stepthe result picks the next step
Fence the steps and budgetit must know when to stop
thinktoolresultadjustdone ✓when all is doneERRANDS✓ flight✓ hotel✓ meetingsnow: packing listone ✓ per stepthink, tool, result, think again — done when it is all done

Think, tool, result, think again. Done when it is all done.Every lap adds one ✓ to the errand list. Tool results pile up on the tray, so more laps mean a heavier tray.

more steps, more chances to slip×90% per step is about 35% over ten stepsif it never stops, it just eats beansa fence: a cap on steps and beansa person checks the risky stepsfixed order? a checklist is cheaperfence the step count and bean budget, and let a person see the end

More steps, more chances to slip. If it never stops, it just eats beans.So a fence caps the steps and the bean budget, and steps like money or delete get a person's check. If the order never changes, an errand checklist is cheaper and safer. Using several parrots: the parrot meeting.

In one breath

Agent = a parrot that, given a goal, plans its own errands, uses the toolbox, and picks the next step from each result. More steps mean more chances to slip, so it needs a fence and a person's check.

A model that takes a goal, plans, acts through tool calls, observes the result and decides again — the ReAct loop. Unlike a workflow with a fixed order, the model picks the next step. Errors compound per step, so you cap steps and cost, add guardrails, and keep a human in the loop.

When grown-ups say it

Agent
The parrot that plans its own errands. Give it a goal and it picks the steps.
ReAct loop
Think, tool, result, think again. One errand per lap.
Planning
Writing the errand list. Breaks the big goal into small steps, and rewrites them as results come in.
Tools
The toolbox. The parrot asks by note and the trainer runs it. → the toolbox
Levels of autonomy
How much you hand over. From asking at every step to going all the way alone. The riskier, the less you hand over.
Agent vs. workflow
A checklist vs. its own plan. If the order never changes, a checklist is cheaper and safer. → the errand checklist
Multi-agent
Several parrots. Split the roles, one parrot coordinates. → the parrot meeting
Guardrail
The fence. Caps on steps, bean budget and forbidden moves. → the fence
Next story
The Toolbox Standard →
Want another term as a picture book? Request →