now, you
are the agent,
and the computer
is your group.
image: ai-generated (gpt-image-2.5-flare)

seven roles

four kinds of cards

the rules

the model sees only its stack, and nobody talks to it.
the harness carries every card to the right tool and brings the result back.
a tool does exactly what the card says, nothing more.
one card at a time, until the model writes ANSWER:.

the system card

search("words")
finds web pages and returns their titles
open("title")
returns the full text of one web page
list_files()
returns the names of all farm files
read_file("name")
returns the text of one farm file
calculator("expression")
calculates, e.g. calculator("4.5 * 100")
write_file("name", "text")
saves a text as a new farm file
this is everything the model knows about its tools.

task 1 · about 5 minutes

how large is the field long acre, and what grows on it this year?

task 2 · swap roles · about 12 minutes

how much nitrogen may we still apply to long acre this spring? give the total in kg N for the whole field.

task 2 · compare

which nitrogen requirement did your model use, and where did it come from?

task 3 · if there is time

plan the first nitrogen application on long acre: how much of which product do we need, and can we spread it on 10 february?
A table during the game: the card for Task 2, handwritten tool and result cards, the system card and stacks of blank cards
every card
on this table
is one step
of the agent.
photo: a group playing task 2

from the game to a real agent

the stack of cards
→
context window: everything the model knows
the system card
→
system prompt: who it is, which tools it has
THINK:
→
reasoning, the plan
TOOL:
→
tool call, written as text
the harness
→
harness: the program that runs the tools
RESULT:
→
observation, added to the context
ANSWER:
→
stop: the goal is reached

the same, in a real agent

TOOL
list the files in the folder Get-ChildItem -Recurse
RESULT
one file: a previous group's report … Applications of Artificial Intelligence.md
TOOL
read that file Read("… .md")
RESULT
“a brief history of artificial intelligence … 1950: alan turing …”
THINK
“before the plan, two things are missing: there's no pinboard photo, and there's no file called milestones.md.”
TOOL
start eight sub-agents Agent("G1 foundations 1943–1969") …
from the agent's log in the experiment “a brief history of AI”; shortened

key takeaway

an AI agent is a large language model (LLM) that can only write text; everything it does in the world happens through tools that the harness runs for it.

reflect

in which round would the model have been stuck without the tools?
what would have changed if the model had to give up its stack after ten cards?
who did the thinking, who did the acting, and which of it could the model actually do?