Architecture

How it works.

Dagny is a chain of specialized agents passing a single typed data structure to each other. One agent, one task, one responsibility. Each step is validated before the next one runs.

The pipeline

  1. 01

    Intake

    You describe your business in chat at bydagny.com/new. Dagny captures what's needed to seed the rest of the pipeline.

  2. 02

    Enrichment

    Three agents run in parallel: scrape the existing site, pull Google Business data, discover social profiles. Results flow back into the same ContextGraph.

  3. 03

    Design & plan

    Design Director picks the visual language. Planner generates a complete WebsitePlan. Reviewer validates it. Founder Gate 1 asks: is this worth shipping?

  4. 04

    Build

    Copywriter drafts every headline, body, CTA, and FAQ. Builder generates a Next.js 16 app. SEO and Agent-Readiness injectors layer on metadata, structured data, llms.txt, MCP, and .well-known.

  5. 05

    Deploy & audit

    Deployer pushes to GitHub and Vercel. Design Auditor checks visuals across viewports. QA Auditor runs Lighthouse and link checks. Founder Gate 2 is the final taste check.

  6. 06

    Live, then maintained

    Polisher applies Gate 2 must-fixes. Site goes live. After launch, the editor accepts customer change requests over email and ships them behind a human gate.

The data

Four typed structures flow through the pipeline. Every field is wrapped in Sourced<T>, which tracks where the value came from and how confident we are in it.

  • ContextGraph — everything we know about the business
  • DesignBrief — emotional target, materials, color hierarchy
  • WebsitePlan — pages, sections, theme, SEO
  • CopyDraft — every headline, body, CTA, FAQ

The stack

LayerWhatRole
AppNext.js 16 customer siteWhat the visitor sees
Pipeline21 agents in agents/*.mdBuild, audit, ship
DataContextGraph → DesignBrief → WebsitePlan → CopyDraftTyped Zod schemas at every handoff
InfraVercel + GitHub + Resend + SupabaseHosting, CI, email, auth

Two human gates

Gate 1 runs after the plan, before any code is written. It asks: is this worth building? If not, planning loops.

Gate 2 runs after deploy and audit, before the site goes live. It asks: is this visibly good? Must-fixes are bundled into a polish pass; only then is the site shipped.