AI-First Engineering DIGITAX ENGINEERING DAY · AI SHOWCASE What changed when I started writing intent and reviewing code. King'ori · A personal perspective from Zappi · 14th July 2026 

QUICK POLL How do you use AI today? Hands up: AI reviews code you wrote Hands up: AI writes code you review Hands up: you shipped agent-written code to production this month 

THE MENTAL MODEL I USE Four years, four eras 2023 2024 2025 2026 Autocomplete. Line-by-line suggestions in your editor Chat. Copy-paste between a browser and your codebase Agents. Tools that read the repo, run the tests, open the PR Orchestration. Engineers directing several agents; whole workflows delegated 

MY ZAPPI JOURNEY · JUN 2025–TODAY Jun 25 Dec 25 Jan 26 Mar 26 Today First AI Engineering Day. I saw what agents might change, but I had not changed my workflow yet Opus 4.5 and the holiday break. I caught up with the online discourse and tried the new tools properly I changed my workflow. I began writing intent and reviewing agent-written code Coding Agents Day. The wider engineering team began making the same shift This is now how I work. Most tasks start with intent, context and review 

FOUR STAGES How my workflow changed CHATGPT I ask questions → AI returns research + snippets INLINE CHAT I direct the change → AI edits inline → I ask and check JAN 26 · AGENTS I write the intent → AI writes the code → I review it LATER · AI REVIEW I write the intent → Build agent writes → → I copy, adapt and write Review agent checks → I review it 

WHERE IT CHANGED HOW I WORK Five areas 1 · Development 2 · Review 4 · Design 5 · Economics Writing intent instead of starting with code Turning ideas into working artefacts Starting review from an agent's briefing Tools I wouldn't have built now exist 3 · Debugging Agents connected to the tools I use 

1 · DEVELOPMENT The unit of work moved from code to intent The brief This presentation “Initialise this as a git repo and commit these changes. Slice them up as you like.” One sentence became six scoped commits: a deck, speaker notes, a demo and a handover kit. Reviewable output One idea per commit 3b5e9cc Add presentation deck and speaker notes 361e777 Add Coding Agents Day playbook e524a63 Add audience task menu e3b4616 Add channel insights story bank f010bc3 Add review-bot reference 090d017 Add automated review starter kit 

2 · REVIEW Useful review depends on context outside the diff Review example repository_sync.py + for repo in repositories: + client.get_metadata(repo) + client.get_permissions(repo) + client.get_status(repo) AI reviewer · first-pass review 1 issue This multiplies external API calls inside a concurrent job. At the configured worker concurrency, three calls per repository can exhaust the provider rate limit and stall unrelated work. Fetch in one batch or bound the concurrency explicitly. Why it noticed: REVIEW.md supplied the operational context the diff alone did not contain. 

3 · DEBUGGING The agent joined the incident, then made the fix permanent 01 Alert 02 Evidence Certificate validation failures The agent follows logs and surface during a live incident. changing system behaviour with the team. 03 04 It correlates the failure with configuration and recent changes. The manual recovery becomes a reviewed code change, not tribal memory. Diagnosis “It felt like an extra team-member.” A real incident. The agent proposed the diagnosis; the team confirmed it. Permanent fix 

4 · DESIGN I made this deck from a brief and a design system Brief Human direction Dark, spare and technical. Make the shift from AI consumer to orchestrator the spine. The model produced the first version. I chose the direction and revised it. WORKING ARTEFACT AI-First Engineering 

5 · ECONOMICS Small internal tools became worth testing I wouldn't have proposed it The coordination and implementation cost I can test it this afternoon I built the internal tool instead of leaving the BEFORE NOW outweighed the value of the idea. idea in the backlog. 

LIMITS AND TRADE-OFFS What got harder The agent's output is your responsibility. Review becomes more important, not less. Moving faster creates a new kind of cognitive load: “AI brain fry.” 

DIGITAX Four questions I'd ask 01 · INTEGRATION 02 · COMPLIANCE Public docs → working sandbox client Domain knowledge → review context Could an agent build from the docs alone? 03 · OPERATIONS Which repeated tasks have clear outcomes? Support · onboarding · rejection triage Which rules live only in people's heads? 04 · DELIVERY Where does engineering work wait? Idea → PR → merge → production 

IF ANY OF THIS RESONATES Three experiments I'd be curious to try 01 · Delegate Give one bounded, real engineering task to an agent and review the result. 02 · Review Trial first-pass review on a non-critical repository and tune what matters. 03 · Learn together Run a Coding Agents Day and let the team decide what deserves adoption. 

TRY IT ON A SAFE REPOSITORY First-pass review on a real pull request. 01 · Add context Define what the reviewer should catch and what it should ignore. 02 · Run 03 · Calibrate Trigger it on a bounded, non- Compare it with human critical pull request. review and tune the rules. 

OVER TO YOU I spend less time producing code and more time specifying, reviewing and deciding. Questions, arguments and scepticism all welcome. Or ask me: what would you delegate first? Where would an agent slow you down? How should review evolve here?