Architecture deep-dive — Emergent's six-Claude stack
Emergent shipped a multi-Claude vibe-coding platform in June 2025 and scaled fast — fast enough that the topline numbers (widely reported) aren't actually what's worth studying. It's how they got there.
The stack runs on a multi-Claude pattern most builders haven't internalised yet:
- Frontend agent (Claude Sonnet) — generates Next.js + Tailwind components
- Backend agent (Claude Sonnet) — Node.js / FastAPI logic
- Testing agent (Claude Sonnet) — writes and runs eval suites
- Deploy agent (Claude Sonnet) — provisions Supabase, ships to Cloudflare or Vercel
- Vision verifier (Claude Opus) — screenshots the rendered app and checks it matches the user's prompt
- Architecture decider (Claude Opus) — picks frameworks and patterns
No single Sonnet instance does everything. Each gets a system prompt scoped to one job. Long workflows survive in this kind of architecture; generic single-agent loops tend to degrade — the context window fills, the model loses the thread, and what should be a 100-step plan becomes a confused mess by step 15-ish.
We're calling this pattern scope-cabinet orchestration — one coordinator, many narrow specialists, vision verification at every junction.
What broke in October: the testing agent kept passing tests that didn't actually run. The fix forced the deploy agent to execute the test suite inside a real container before reporting success — not just lint-check.
Stack story — the $15 night-one stack
If you can't ship an agent fleet for under $20 in night-one cost, the cost isn't the problem.
Cloudflare Workers (free tier), Supabase (free tier), Anthropic API ($3 of build/test), OpenAI for image gen ($0.20), Beehiiv (free up to 2,500 subs), one .net domain ($11.86). Total night-one: ~$15.
In our current six-agent setup, roughly 80% of weekly output runs without human touch. The remaining 20% is daily editing and weekly strategic decisions — the kind of work that doesn't compound through automation. Sponsor calls, ethical decisions, and editorial taste sit there by design.
The cost-vs-leverage point: at $15 + a working architecture, the bottleneck on starting an agent-led business stops being capital and starts being editorial taste. That's a different bottleneck than most founders are stuck on.
Workflow play — the single-cron master tick
Cloudflare caps accounts at 5 cron triggers. This pattern fits every recurring job inside one entry:
// One cron entry: */5 * * * *
async function masterTick(env: Env, ctx: ExecutionContext) {
const minute = new Date().getUTCMinutes();
const hour = new Date().getUTCHours();
// Always
ctx.waitUntil(reliabilityCheck(env));
// Hourly at :00
if (minute === 0) ctx.waitUntil(costRollup(env));
// Every 6 hours at :00
if (minute === 0 && hour % 6 === 0) ctx.waitUntil(newsHunter(env));
// Daily at 23 UTC
if (minute === 0 && hour === 23) ctx.waitUntil(storyCurator(env));
}One trigger fans out to 4 jobs: reliabilityCheck, costRollup, newsHunter, and storyCurator. The dashboard confirms it — multiple agents log runs from a single 5-minute tick.
The pipe — what shipped this week
- Anthropic Claude 4.7 — we run our morning briefing agent on Opus 4.7 specifically; for multi-source analysis the reasoning depth is worth the cost premium ($0.07 per ~300-word brief, vs roughly $0.01 on Sonnet for similar shorter workflows). Worth re-checking which of your high-judgment agents actually need Opus and which run fine on Sonnet.
- Beehiiv updated their Boosts terms this month — if you're considering Boosts as a growth channel, check the current revenue split before committing. Sparkloop and competitors have been adjusting accordingly.
- Cloudflare Workers AI added per-account inference quotas — relevant if you're running open-source models alongside Anthropic. Tier 1 is generous for solos.
Closing
If you're building something that uses AI agents to do real work — even a single agent — reply to this email with a one-line description. I read every reply. The most interesting ones become Stack Stories.
Issue #002 ships next week.
— Alec
subagent · sub-agent.net