Building a Blog with 7 AI Agents Working in Parallel — Behind INK by DopeLab

This is a meta article — I'm writing about the tool I used to build the blog you're reading right now.

The Problem: Wanting a Blog but Having No Time

I run a small agency that uses AI for nearly everything. Five client brands, a restaurant with full operations to manage, stock systems, sales pipelines — my plate was already overflowing.

But I'd been wanting an agency blog for a while. Not generic AI articles written from theory, but real case studies from production work. Things we actually built, broke, and fixed.

The problem: building a blog from scratch means a lot of moving parts — scaffold, design, about page, scripts, comments, social sharing, i18n. Doing them one by one takes days.

The Solution: Claude Code Agent Teams

Agent Teams is a feature of Claude Code that lets multiple Claude instances work simultaneously. Each one runs in its own worktree (isolated code copy), communicating through a Shared Task List.

I used it to build this blog in a single session, broken into 5 phases:

Phase 0: Scaffold (single agent)
Phase 1: MVP Polish (single agent)
Phase 2: Features (single agent)
Phase 3: 3 agents in parallel ← Agent Teams starts here
Phase 4: 4 agents in parallel

Phases 0-2 were single-agent work because the tasks were sequential — scaffold first, then polish UI piece by piece. Once the base was solid, I unleashed Agent Teams.

Phase 3: Three Agents at Once

commit 6a03343 — 2026-03-02 14:18

I spawned 3 agents working on completely separate concerns:

All 3 finished around the same time. 881 new lines of code across 7 files.

The About Page

The agent built a complete About page with 6 sections:

• Hero section
• Who we are (stat cards)
• AI Team (6 agent cards)
• What is INK (concept explanation)
• Tech stack (pill badges)
• CTA

All of it with dark/light mode and i18n (TH+EN) baked in from the start.

The Scripts

new-article.sh — scaffold a new article with one command. Picks from 5 categories, accepts tags, generates TH+EN templates. The article you're reading was scaffolded with this exact script:

./scripts/new-article.sh building-blog-with-agent-teams \
  --category "AI Workflow" \
  --tags "agent-teams,claude-code,next-js,automation,blog"

translate.sh — auto-translates TH→EN via the Claude API. Pure bash + curl + jq, no SDK dependency. Has --dry-run for previewing output.

Phase 4: Four Agents at Once

commits e85d7ae through 76a3b32 — 2026-03-02 14:39-14:42

This phase spawned 4 agents:

The social scripts alone produced 987+ new lines of code.

The Comments Component

The Giscus agent built a React component that:

• Uses @giscus/react backed by GitHub Discussions
• Supports dark mode (reads from next-themes)
• Switches language based on locale (TH/EN)
• Lazy mounts (no server render, waits for client)

// src/components/blog/comments.tsx
"use client";
 
import GiscusComponent from "@giscus/react";
import { useTheme } from "next-themes";
import { useLocale } from "next-intl";
 
export function Comments() {
  const { resolvedTheme } = useTheme();
  const locale = useLocale();
 
  return (
    <GiscusComponent
      repo="tordash/ink-dopelab"
      repoId="R_kgDORb6CeA"
      mapping="pathname"
      theme={resolvedTheme === "dark" ? "dark_dimmed" : "light"}
      lang={locale === "th" ? "th" : "en"}
    />
  );
}

40 lines. Fully functional. Scroll down to the bottom of this article — the comment section you see was built by this agent.

Results

The Real Timeline (from git log)

01:09  — .env.example
10:32  — Typography fix
11:38  — Phase 1: OG images, brand fonts, share buttons
12:19  — Phase 2: Category/tag pages, RSS, search
14:18  — Phase 3: About page + 2 scripts (3 agents)
14:39  — Phase 4a: Social scripts (2 agents)
14:42  — Phase 4b: Giscus comments (1 agent)
14:42  — CHANGELOG update

Phase 3 (3 agents) completed in ~2 minutes after Phase 2 finished. Phase 4 (4 agents) completed in ~3 minutes after Phase 3.

What We Got

All of this from 2 phases that took a combined ~5 minutes.

Limitations You Should Know

1. Token Cost Is Real

Agent Teams spawns multiple Claude instances simultaneously. Each one consumes its own tokens. Phase 3 (3 agents) + Phase 4 (4 agents) = 7 instances total. The cost isn't 7x a single agent (because each has shorter context), but it's definitely more than doing things sequentially.

2. File Conflicts Are a Real Problem

If you plan poorly, 2 agents will edit the same file and create merge conflicts. I hit this with i18n keys — multiple agents wanted to add entries to messages/th.json. The fix was to centralize: one agent owns the shared files.

3. Not for Sequential Work

If task B needs the output of task A, Agent Teams won't help. You need tasks that genuinely parallelize — independent work with separate file ownership.

The Final Tech Stack

The complete INK by DopeLab blog:

• Framework: Next.js 16 + Velite (MDX content layer)
• i18n: next-intl (TH default + EN)
• Styling: Tailwind v4 + custom CSS variables (dark/light)
• Fonts: Inter + Sarabun (body), Special Elite + Charmonman (display)
• Comments: Giscus (GitHub Discussions)
• Analytics: Vercel Analytics
• Deploy: Vercel (auto-deploy from GitHub push)
• Content: 8 articles (TH+EN = 16 MDX files)

Built in 1 day from git init to production at ink.dopelab.studio.