9
specialised agents
← work
Case study · platform · 2026
email innovation hub: nine agents behind a quality gate.
A platform where AI agents scaffold, fix and review production email, and a ten-point automated gate decides what ships. Designed and built solo, on my own time, around the hardest consistency problem I know.
The context
Ten years inside agency email teams showed me the same day, everywhere: a comp rebuilt by hand in table HTML, rendering in the browser and breaking in Outlook, dark mode inverting a logo, QA squinting at ninety Litmus screenshots at 6pm. The expensive part was never the first version; it was holding one design consistent across dozens of rendering engines.
The bet
If the consistency rules can be written down, machines can hold them. Agents do the scaffolding, dark-mode surgery, Outlook fixes and copy variants; a ten-point automated gate checks everything before a human reviews. People keep the brief and the brand.
The design decisions
The workspace is a single surface.
Code editor, live preview and agent chat sit together, because email work is a conversation between intent and rendering, and every tab switch between those was waste.
The gate is a UX, not a report.
Each of the ten checks returns a verdict a non-developer can act on, next to the thing it refers to, so a failed check becomes a task rather than a finding.
The design system is built from the tokens up.
Tiered tokens, primitives beneath a semantic contract, atomic components above, one icon sprite, one shell. Agents generated the platform's densest surfaces (workspace, QA engine, approvals, render views) against that single contract, and they read as one product. Because components touch only the semantic tier, the system re-skins in one file.
Design syncs from Figma.
The design source of truth stays where designers work; the platform pulls it into coded, versioned components, so the system in Figma and the system in production cannot quietly diverge.
Approval is part of the product.
Clients review, comment and sign off inside the platform with an audit trail, because in agency life the approval loop is where weeks actually go.
Personas make QA honest.
Test profiles for device, email client and dark mode, so "does it work" is a checklist run by machines, not a hope.
The platform remembers.
A retrieval-backed knowledge module, pgvector with hybrid search, answers from the team's own accumulated craft: past fixes, client rules, rendering quirks and the decisions behind them. Agents consult it before proposing, so the platform's judgement compounds instead of resetting with every brief. This is the design knowledgebase from the practice page, in production.
What shipped
The MVP is deployed and live: FastAPI and Next.js, 261 endpoints, 7,055 tests, 867 commits, designed and built solo on my own time, with the IP mine. Several of its features were put forward for adoption at Dentsu before my role was made redundant.
What I learned
The gate outperforms the most careful reviewer at 6pm, and it never gets tired of Outlook. Quality held by machines, so people spend their taste where it counts, is now the centre of my practice.