Managed Agentic Engineering · reference implementation
How this site was built with MAE
Spec-driven development where a person defines what and why in structured documents, while an AI agent executes how under explicit gates. Below is the complete flow, real figures, and honest limits on their accuracy.
· repository git history + Discovery artifacts
TL;DR — key figures
Work traditionally estimated at 26–40 person-days took about 25 hours in the commit-based reconstruction. This is not a controlled benchmark: the method and every limitation are shown below.
Method
What MAE is — in two sentences
MAE is supervised agentic development: routing, planning, and execution complexity lives inside a pipeline, while a person reviews decisions at explicit gates. Repository documentation is the source of truth.
- 1PCDProject Context
- 2FSDepics → features → US + AC
- 3spec.mdbuildable Feature Spec
- 4plan.mdAI Implementation Plan
- 5Gate 3engineer approves the plan
- 6Executionagent follows the checklist
- 7Quality gatelint · typecheck · test · build
- 8Docsfeature record + map
- 9Merge → deployhuman approval and release
Phase 1 · 24–27 July 2026
Discovery — resolve uncertainty before the first line of code
Auditing the old Gatsby site produced facts that changed the initial assumptions.
- 7 articles; 5 hosted on otakoyi.software
- Migration became a small task and External Post became a distinct entity.
- Last publication: November 2022
- Content cadence—not technology—became the number-one risk.
- 4 categories + 8 legacy tags
- The URL taxonomy was preserved through an explicit compatibility map.
- No analytics
- analytics-and-seo was added because the main metric was otherwise unmeasurable.
- English-only content
- Full bilingual delivery became a strategic decision.
28 decisions, 11 risks
Discovery recorded 28 decisions across three iterations. The D12–D13 pivot made OTAKOYI lead generation an end goal, moving Contact, GDPR, and Privacy into M1.
R9 stayed open deliberately: developer traffic does not guarantee CTO enquiries. Plausible data will test the hypothesis; review is scheduled for 2027-03-31.
Phase 2
Specification — Feature Index → spec.md
One Feature Index row equals one feature, one spec.md, and one branch. The slug travels from Discovery to code; an acyclic dependency graph controls build order.
Foundation
M1- i18n-routing
- analytics-and-seo
Personal Brand
M1- home-landing → home-conversion-ux
- about-page
Content & Blog
M1 · M2- blog-index-and-post
- content-categories
- content-tags
- content-search
Lead Generation
M1- contact-lead-form
- privacy-policy
Audience Growth
M2- newsletter-subscription
Exactly 11 Feature Index entries; content-search and newsletter-subscription were deliberately deferred to M2.
Phase 3 · 28 July – 11 August
Build — the real git trail
Next.js 16 · React 19 · strict TypeScript · Tailwind v4 · MDX · Resend · Upstash Redis · Plausible · Vercel. No database; the only write path is outbound email.
Discovery
Audit, 28 decisions, 11 risks, and Q3 OKRs.
Bootstrap — /mae:init
Next.js base, MAE SDD layer, PROJECT.md, and architecture map.
Core build
Foundation → Content → Lead Generation by dependency graph.
Consolidation and infrastructure
256 files, +16,009 / −1,279 lines; CI, Vercel, and legacy content.
M1 formally closed
Canonical www, D27 articles, article template, and production verification.
Post-M1 polish
Conversion UX, About refresh, mobile navigation, and localized CV.
What “M1 closed” means
- 47 canonical URLs returned 200
- 6 normalization cases returned a direct 301
- 3 D27 articles returned 200 in their declared locale
- The sitemap index reported Success in Search Console
- A control lead was delivered with the correct Reply-To
- HTTPS/HSTS, DKIM/SPF/MX, Upstash, and two Plausible goals were verified
Effort
How much time it took
Git shows 47 commits across six active days between 28 July and 10 August. This is a reconstruction, not an automatic time tracker.
Estimation method
Commits were grouped into sessions: 11–16 hours between commits, 15–23 hours after adjusting for session starts, plus a 5–8 hour blind spot before the consolidated commit. The recommended range is 20–30 hours, centred on ≈25.
Estimate vs actual
Traditional M1 estimate: 26–40 person-days
Commit-based reconstruction: ≈3–4 working days
Indicative compression: 8–13×Human + agent
What the agent could do—and what it could not
The contact-lead-form sample shows that an agent drafts what follows mechanically from Discovery; component states and API contracts still require human decisions.
Quality control
Why it works—and why fast does not mean careless
Speed comes from resolving uncertainty early and repeating verification gates, not from skipping steps.
Discovery resolves risks
8 of 11 risks closed before the first line of code.
Concrete specifications
Numbers, limits, TTLs, and error codes exist before execution.
Quality gate
lint → typecheck → test → build; releases cannot bypass it.
Gate 3
A human hard stop between plan and execution.
Docs as source of truth
The pipeline maintains feature and fix records.
Production verification
47×200, 6×301, delivered lead, and sitemap Success—not assumptions.
What this means for OTAKOYI clients
MAE turns an uncertain digital initiative into a managed decision system: Discovery, a buildable spec, controlled execution, and evidence-based verification.
Figure sources
- Git history: 59 commits as of 2026-08-11; effort snapshot: 47 commits.
- requirements/discovery-summary-maluk-tech.md — D1–D28 and Risk Closure Log.
- requirements/project-context-document-maluk-tech.md — scope and traditional estimate.
- requirements/m1-release-reconstruction-contract.md and m1-production-url-audit.md.
- requirements/feature-specification-document-maluk-tech.md and features/*/spec.md.
- requirements/okr-q3-2026.md.