Stop typing prompts. Build the loop.

02 · the page

Everything it knows fits on one page.

The model remembers nothing between calls. So each turn, the code around it (called the harness) writes it a fresh page: rules, task, tools, and everything that's happened so far. That page is the context.

TakeawayIf it isn't on the page, it doesn't exist.

03 · the model

It reads the page and writes one next step.

The model is the one part you rent instead of build. It can't run anything. It reads, and writes text back. This time the text is a plan: see it fail first.

PinnedAnthropic: agents are "typically just LLMs using tools based on environmental feedback in a loop."

04 · the tool call

A tool call is a request, not an action.

It picked from a menu you wrote: three tools, each with a name and a shape. It asked to run the tests. Nothing has happened yet. Your code decides whether, and how, to do it.

TakeawayA few clear tools beat many vague ones.

05 · the world

Your code does the doing.

The harness runs the tests inside a sandbox, a walled-off copy of the project, and captures what comes back, exactly as it came back. It's the only moment in the turn that touches anything real.

TakeawayThe result is what the model reads next. Keep it short and true.

06 · the check

Done, or go again?

A machine makes this call, not the model. All tests green: stop and hand back the answer. Otherwise go again, unless the budget is spent. Then the loop stops and a person takes over.

PinnedThe research review: runs stop "when a machine-checkable condition holds."

The run sheet

Everything you just watched is one line.

Trigger, page, model, tool call, world, check: that was one turn. The harness writes each turn onto the run sheet as one line. Watch the end of the line.

The carriage return

What happened goes on the next line.

Think of an old typewriter. At the end of a line, the carriage swings back to the left margin, one line down. Here, the test result rides that return onto line 02's page. That's the loop.

▸ the same idea, as LangChain puts it

LangChain · The Art of Loop Engineering · Jun 16, 2026

Line 02 · it types by itself

Now nobody's pressing keys.

Line 02's page includes the failure, so the model asks to read the code. It finds the maths: each share is rounded to the cent, so three shares of $33.33 add up to $99.99. A cent goes missing.

Line 03 · a fix

A fix that looks right.

Give the last person whatever's left over. On paper, that's perfect. And because the rules say so, the harness reruns the tests after every edit.

Line 03 · the return

The failure rides back down.

That exact line, 23.340000000000003 != 23.34, is carried to the left margin and lands on line 04's page. The next fix will be written with it in view.

Line 04 · the fix

Count in whole cents.

So the next fix is different: do the maths in whole cents, where nothing can round wrong, and give the leftover cent to the last person.

+    cents = round(total * 100)+    base, extra = divmod(cents, people)+    shares[-1] += extra$ pytest -q   4 passed in 0.03s

Exit

It leaves with its evidence.

Four tests pass, so the check says done. The answer goes out through the right margin with the change and the test run attached, so a person can see why it's trusted.

After line 04

Four lines. Two to spare.

One bad fix, caught by the loop itself, and two lines of budget left over. The sheet is also a record: every page, call and result is still there for a person, or the next loop, to read.

Illustrative runEvery tool output is real (pytest 9.1.1, Python 3.14, 27 Sep 2026). The model's choices were written for this page.

The loop above

And there's a loop above this one.

Finished sheets are a record of what worked and what didn't. LangChain describes reading them to improve the rules, the tools and the check for the next run, and calls it a hill-climbing loop. The outbox feeds the page.

Where it started

One line of Bash is already a loop.

Geoffrey Huntley's "Ralph" runs the agent on the same prompt file, forever. Everything else on this desk is what you add so it knows when it's done, when to stop, and who to hand it to.