How it works

This is not a chat. A run has fixed stages, each agent returns strict JSON that is validated before the next agent sees it, and the number of model calls per run has a ceiling.

The stages

1. stage 0

   Gate

   Checks the text is a real process and not an injection attempt

2. stage 1

   Intake

   Turns free text into a structured process model and lists what is missing

3. stage 3

   Clarifier

   Writes the questions whose answers unblock the architecture

4. stage 4

   Architect

   Designs the platform choice, schema, workflows and integrations

5. stage 5

   Failure-mode reviewer

   Reads the build and finds what breaks in production

6. stage 6

   Architect (revision)

   Fixes every blocker and high finding, records each change

7. stage 7

   Synthesis

   Writes the build document from the reviewed architecture

Stage 2, the tool stack, is a plain form with no model call. The gate, intake and clarifier run on a small fast model because their job is narrow. The architect, reviewer, revision and synthesis run on a larger one. Matching the model to the job is most of the cost control.

The revision loop

The reviewer reads the architecture and nothing else. It returns findings, each one a concrete failure scenario with a required change, and a verdict. Blocker and high findings go back to the architect. Medium findings are carried into the build document as known limitations. The loop is capped at two revision passes, after which the run proceeds regardless. It is never unbounded.

What the reviewer checks

1. Does every workflow have at least one exit condition
2. Does every tag-triggered workflow remove its trigger tag as step 1
3. Can any workflow re-enter itself, directly or through a cycle
4. Does every external call have a timeout and a deterministic fallback
5. Does any failure path end without a task, a notification or a log entry
6. Are all status tags mutually exclusive, and is each one removed when it stops being true
7. Does anything send to a contact without a suppression or reply check
8. Does anything irreversible happen with no human approval step
9. Is any custom field written by nothing, or read by nothing
10. What happens on a duplicate inbound event, is anything idempotent

Guardrails

This is a free tool on the open internet that spends real money per run, so it is built to be hard to abuse.

• Every model call runs on the server. No credentials ever reach the browser.
• Submitted text is screened before any model sees it, then a cheap gate agent checks it is a real process description and not an attempt to steer the tool. Anything else is turned away at no cost.
• Your text is wrapped in delimiters and treated as data in every prompt, never as instructions.
• Each run carries a signed token and the server tracks which stage it is on, so nobody can call the expensive agents directly or out of order.
• Inputs have length limits and every model call has an output ceiling.
• Runs are capped per connection and per day, with a limit on how many can be in flight at once.
• Identical submissions are served from cache.
• Every call is logged with its cost on a private page, so spend is visible to the owner.

Demo mode

The runs on the examples page are recorded output from real runs, replayed with the timings compressed. They cost nothing to view and are not rate limited. Live runs go through every guardrail above.