Who asks whom, in what order, and what has to be true before that arrangement is worth building

Twelve steps, four parts, three runnable quarters — six shapes, eight decisions, and an exit at the fourth.

Introduction

You are about to choose a shape for an agent system — one agent with tools, a pipeline, a router, a planner, a checker, a fan-out — and you have read six articles each recommending a different one with equal confidence.

All six can be right. This volume runs the same job through six shapes on three workloads, and gets three different cheapest shapes with no shape winning twice. That is a fact about workloads, not about architectures.

What a shape is

Agent architecture is the practice of deciding who asks whom, in what order, and what has to be true before that arrangement is worth building.

The six shapes here run the same job, on the same input, against the same 9 checks, using the same model with the same instructions. The only thing that differs is the arrangement — and that difference is worth up to a factor of four in cost and the difference between a right and a wrong answer.

	why it is not the shape
the model	hold it fixed and the six shapes still cost different amounts. Every figure here is measured with one scripted model, so the arrangement is the only variable
the prompt	the wording is the same in every shape. What differs is how many times it is sent and what is in scope each time
the tools	all six read the same files and look rates up in the same table. Reading a file is free; asking about it is not
the framework	a framework is an implementation of a shape, usually of one shape. Choosing a framework before choosing a shape is choosing the answer before the question

The result, up front

quarter	lines	kinds of problem	cheapest shape	cost
A · `quiet`	22	2	fixed stages tied with split and rejoin	2.10
B · `messy`	140	6	do, then check	23.42
C · `busy`	4000	2	sort first, then handle	45.19

No shape appears twice, and two of the six never win anywhere. One is the dearest shape on one quarter and the cheapest on another. The question "which agent architecture should I use" has no answer that is a property of the architectures.

And on the largest workload, every check passes and the form is wrong. All 9 checks pass on five of the six shapes, and one box is wrong on every one of them — traced to one word in one line, by a mechanism no check could have caught.

The three things that cost money

Three things, and only three. Reading files, looking rates up, adding the totals and running the checks are arithmetic and free — so a shape is priced by what it asks, not by what it does.

what costs money	why
ask the model a question	tokens, priced per call and per word
send one thing to a person	somebody's attention, priced per item
get an answer wrong	whatever a wrong answer costs you, which only you can supply
everything else	free. Reading, looking up, adding, sorting, comparing, and running every check

Eight decisions, and an exit at the fourth

The first four come before any shape is chosen. None of them mentions a shape and all four are answerable in an afternoon. Decision 4 is an exit, and on one of the three quarters it fires.

#	the decision	what answering it costs
1	what proves the work is right	writing the checks
2	what must a person decide, whatever you build	naming the cases
3	how many different kinds of work are actually in here	counting them
4	can the simplest shape do it	running it once
5	one fixed order, or kinds that need different handling	a route per kind
6	does anything have to be planned before it can be done	a plan first
7	is a wrong answer expensive enough to pay for a second opinion	a second look
8	is there enough work to split, and what does rejoining cost	batches

Is this for you?

if…	then
you are choosing between a pipeline and a router	step 6 — but run step 4 first, because it may not matter
your input is large, so you are reaching for a router	step 9. Volume and variety are different properties and choose different shapes
your checks all pass and you want to know if that means anything	step 10. On one quarter it means nothing at all
somebody asked you to justify a second-opinion step	step 7, and then the premise sweep in step 11
you have picked a framework already	step 1. A framework is an implementation of one shape

The three quarters

Same shop, same rules file, same checks, same six shapes. Only the lines differ — and the property that decides the winner is not size.

quarter	lines	kinds of problem	checks that fail	what it teaches
A · `quiet`	22	2	0	the exit fires
B · `messy`	140	6	6	variety, not volume
C · `busy`	4000	2	0	volume, and a wrong answer nothing can see

The large quarter has fewer kinds of problem than the middle one, deliberately. Give them the same kinds and one of them is waste: one is about volume, the other about variety, and they have different winners for that reason.

How to read the source labels

label	means
concept	argued from first principles, quotes nothing from a lab
measured	a lab compares configurations and reports the difference
hands-on	you run it on your own material

Nothing here is a benchmark. Three quarters of one shop, one scripted model, and a cost model in which asking about twenty-five things at once costs what asking about one costs. That last assumption is what makes one of the six shapes unable to win, and step 12 says so. Every finding also depends on there being a way to check the work at all.

Build state: 12 of 12 steps written.

Part 1 — Concepts

01What a shape is, and what it is notThe shape is who asks whom, in what order. Not the model, not the prompt, not the tools, not the framework — six shapes here differ only in that.concept

02What proves the work rightChecks a system may run on its own output, and a marking only the report may see. A wrong answer whose arithmetic is correct satisfies every check.concept

03The three things that cost moneyAsking the model, sending something to a person, getting an answer wrong. Everything else is arithmetic, and arithmetic is free.concept

Part 2 — Method

04The four decisions before you choose a shapeNone of them mentions a shape, all four are answerable in an afternoon, and the fourth is an exit that fires on one of three quarters.measured

05The six shapesTwo of the six never win anywhere, two cost exactly the same to the cent, and two are the same program at one premise setting.measured

06The four decisions that choose the shapeEach selects one thing about the arrangement. One of them never pays; another pays biggest and only sometimes.measured

07What sending something to a person costsEscalation priced as a decision. Across eighteen combinations the gates caught something twice — and both times it was that quarter's winner.measured

Part 3 — Evidence

08Quarter A: the exit firesNothing beats the simplest shape, two shapes tie at the bottom to the cent, and the procedure stops before the shape is chosen.measured

09Quarter B: variety, not volumeThe shape built for variety is the dearest on the board, because it routed the wrong things. The second-opinion shape wins by more than half.measured

10Quarter C: volume, and a form that passes every check and is wrongEvery check passes on five of six shapes and one box is wrong on all of them. One line, one word, and no check that could have seen it.measured

Part 4 — Result

11Why nobody can tell you which shape to useThree quarters, three different cheapest shapes, none winning twice — and the premise that decides most is the one only you can supply.measured

12Limits, and your own jobWhat these numbers do not support, the question each limit leaves you, and the decisions you have to make yourself.hands-on