Which system shape should you build?
Seven questions about the ask, the input, who reviews it, and what fails you can't survive. No model call, no chat, just the same deterministic scoring behind the one-paragraph answer on /architectures/.
Read-only, hands back an answer, a human reviews every one: that shape can often ship as a fixed pipeline with no agent loop at all. Read Agents vs. pipelines →
The bolt-on architecture, the parts teams skip and regret, the cost lever, and the honesty check: enter your email to unlock.
Helps us send you relevant follow-ups. Skip any time.
What this measures#
The picker scores seven questions against six candidate architectures, then returns the top scorer plus a possible second pick. Each answer option carries a weight toward the architectures it actually points at. The result follows the shape of the problem, not which option read best. The first question alone carries the most weight, since it is the most direct statement of what someone wants built. The other six refine that first read rather than overturn it.
The seven questions
Every question does one job: it rules some architectures in and some out. None of them ask about the model or the vendor behind it, because neither choice changes which of the six shapes actually fits the problem. The full wording of each question is shown here, next to what it is really testing for.
| Question | What it asks | What it separates |
|---|---|---|
| The ask | What does someone actually ask this system for? | Points straight at one of the six architectures, based on how the reader describes what they want built. |
| What it returns | What does it hand back? | Separates the six by the shape of the output: a sourced answer reads very differently from a block of merged code. |
| Input source | Where does the input come from? | Separates the six again by where the work starts, from a document set to a connected repository. |
| State changes | Does it change state anywhere? | Places the system on a spectrum. One end never writes; the other writes on its own, with no one asking first. |
| Human position | Where does the human sit? | Places the human on that same spectrum. At one end, a person reviews everything; at the other, no one is watching. |
| Latency & volume | What is the latency and volume shape? | Separates a system that must answer within seconds from one that can run as a scheduled batch. |
| Unsurvivable failure | What is the failure you cannot survive? | Separates the six by which failure would actually sink the project, from a confidently wrong answer to being locked to one vendor. |
The six candidate architectures
Scored in this fixed order. A tie between two architectures breaks toward whichever one is listed first here. The order is pinned deliberately, so the same seven answers always return the same pick.
| Order | Architecture |
|---|---|
| 1 | Grounded answering over your own documents |
| 2 | Agentic workflow runner for a back-office queue |
| 3 | Document extraction with a human review loop |
| 4 | An assistant inside a product you already ship |
| 5 | Model-agnostic inference gateway |
| 6 | A coding-agent delivery fleet |
What this does not do A second architecture surfaces as a likely bolt-on once its score clears 70% of the primary score. Below that line, the picker says plainly that there is no clear second pick. One answer combination triggers an extra note: a read-only system that hands back a sourced answer with every output under human review. When that combination shows up, the picker adds that a fixed pipeline might already be enough, with no agent required. The picker assumes the build matches one of six shapes reasonably well. A build that genuinely straddles two of them still gets a forced-fit answer. So does one that matches none of them. Read that answer as a first draft of the shape. The build plan comes after. Seven questions can narrow a build down; they cannot see the team behind it or the one internal system nobody wants to touch. Those still belong in a real scoping conversation.
The picker is a free, self-serve first pass. A Ship Audit is a fixed-scope, fixed-fee engagement that checks which of these your actual system needs against your real constraints, not a seven-question guess.
Book a Ship Audit →