Chapter 03
The Method
Five specialized agents, one structured pipeline, one belief about how software should be built.
The studio’s multi-agent development method is the reason this catalog is the size it is. Five specialized AI agents work in concert with structured JSON handoffs, confidence scoring, and automatic fix-iteration loops.
The method is an orchestration layer: a set of five explicitly-scoped roles that pass typed artifacts to each other through a JSON handoff protocol, with a Reviewer in the middle that rejects work below a confidence threshold. The result is a development cadence most small studios cannot match.On role disciplineThe hard part is role discipline: keeping agents out of each other’s scope. Each role ships with an explicit refusal list, things it will decline to do regardless of what the surrounding prompt suggests.
Mature framework. Production-tested through the studio’s own development workflow.
The pipeline
Five roles, explicit handoffs.
-
Planner
Decomposes the ticket into an ordered plan. Selects the workflow path and emits a structured brief.
Handoff PLAN.json
-
Context-Fetcher
Pulls the files, schemas, and prior decisions the Feature-Builder will need. Writes a grounded context packet.
Handoff CONTEXT.json
-
Feature-Builder
Writes the code. Scoped to a single feature, operating against the context packet and the plan.
Handoff DIFF.patch + NOTES.md
-
Code-Reviewer
Reviews the diff against the plan. Checks for CVEs, policy violations, and architectural drift. Assigns a confidence score.
Handoff REVIEW.json (score 0–1)
-
Test-Writer
Authors the tests the Reviewer required. Runs them. Iterates until green or until human intervention is flagged.
Handoff TESTS.json + CI result
Why it works
Structured handoffs, confidence scoring, and refusal lists.
The Reviewer emits a numeric confidence score between 0 and 1. Below a configurable threshold the pipeline loops: the Feature-Builder gets a new plan segment, the Reviewer re-scores, the Test-Writer re-runs. Above the threshold the work ships. Every artifact is a JSON document that can be inspected, replayed, and (importantly) diffed across runs to understand how the method evolved a given feature.
The method built this site. Chapter 03 is, in that sense, a demonstration: every page in this manual was planned, contextualized, built, reviewed, and tested by a different role in the same pipeline. The Colophon elucidates which runs produced which chapter, which packages landed in the stack, and where we made substitutions along the way.
Next · Chapter 04
Working Together
Six engagement models, direct contact, no fuss.
Read Chapter 04 →