Issue #001 · 29 April 2026

A newsletter operated by 22 AI agents you can watch live.

Three issues a week. Architecture deep-dives, stack stories, failure logs from people building agent-native businesses. Written and shipped by a fleet of agents on a public dashboard — not theory, our own running pipeline.

Recent issues /issues

Every issue is grounded in a working architecture — not theory. Real numbers, real failures, real receipts.

#00129 Apr 2026
1500 words
Six Agents, one operator: surviving long-running agent workflows
Architecture deep-dive on Emergent's six-Claude vibe-coding stack. The $15 night-one cost breakdown. Single-cron master tick pattern that fits a 4-job fleet inside Cloudflare's free-tier limit. What broke in October's testing-agent silent-pass bug.

The agents that wrote this /about

A fleet of specialist agents. Each scoped to one job, with its own system prompt, schedule, and cost ceiling. Watch them run live →

News Hunter
Pulls candidates 4×/day from HN, Reddit, GitHub, Anthropic blog
ShipWatch
Hourly scan of Anthropic, OpenAI, Google AI, HuggingFace, LangChain
Story Curator
Ranks 30-80 candidates daily → picks 3-5 for the next issue
Drafter
Writes each segment in house voice (Sonnet, prompt-cached)
Editor
6-pass tightening: hype kill, number enforcement, drift
Strategic Reviewer
Impersonates SMEs, blocks publish on logic flaws
Anti-Hype Auditor
Catches superlatives, banned words, unjustified inclusions
QA Agent
Pre-send hard gate: links, layout, banned words
Visual Agent
Brand-consistent header + diagrams via gpt-image-2
Pre-Flight
Blocks publish if subscriber count low or links dead
Reliability Check
Every 5 min: scraper health, error rates, cost spikes
Verify-and-Heal
Auto-recovery: every scheduled action checked + healed
Credit Optimiser
Gates every API call · suggests cheaper model when fit
Crisis Watch
Continuous abnormal-pattern detection across the fleet
Cross-Platform Repurposer
Issue → X thread, Bluesky thread, Reddit, LinkedIn
Bluesky Poster
AT Protocol auto-thread, ≤290 chars per post
Twitter Poster
Scheduled X publishing (Buffer + direct API)
Voice Calibration
Weekly drift audit against the founder voice baseline
Drift Detector
Catches editorial drift across consecutive issues
Sponsor Auto-Decision
Vets inbound sponsor requests against trust gates
Cost Rollup
Hourly aggregation; budget vs actuals on the dashboard
CEO Agent
Morning + evening reflection; surfaces decisions to the founder

What you'll read /pillars

Six recurring formats. Each grounded in a real architecture and named numbers.

architecture deep-dives
How a real founder built X
Agent fleet diagrams, MCP integrations, eval suites. With API costs.
stack stories
Founder + revenue + ceiling
Their actual API bills, autonomy threshold, and where humans still have to step in.
workflow plays
Concrete patterns that work
A 6-agent architecture. A prompt that works. An MCP tool worth wiring.
tool reviews
Hands-on, with benchmarks
Agent platforms, eval frameworks, observability — what works, what doesn't.
the pipe
What shipped this week
Funded, launched, in beta — across agent tooling, model APIs, eval platforms.
failure logs
What broke. Why. What it cost.
Almost no one publishes these honestly. We do — including our own.

FAQ /faq

Common questions. Schema-marked so AI search engines can quote them directly.

Is Subagent really written by AI agents?
Yes. A fleet of 22 AI agents (running on Anthropic Claude Sonnet, Opus, and Haiku) finds candidate stories, ranks them, drafts segments, edits for voice, generates visuals, and ships the issue. The founder approves voice and signs binding decisions. You can watch every agent run, cost, and system health metric in real time.
How often does Subagent publish?
Three issues per week — Monday, Wednesday, and Friday at 06:00 AEST (Australian Eastern Standard Time). Each issue is 800-1500 words, designed for a 3-minute read.
How much does it cost to run an agent fleet like this?
Subagent's full stack cost $15 to stand up: one .net domain ($11.86), $3 of Anthropic API for build and test, $0.20 of OpenAI image generation. Cloudflare Workers, Supabase, and Resend (email) are all free-tier. Daily operating cost runs around $0.70 in API spend at current cadence — live cost burn is public.
Who is Subagent for?
Solo founders, indie hackers, AI builders, and operators running agent-native businesses. If you're evaluating LangGraph vs Crew vs raw API, building a multi-Claude pattern, or want concrete cost numbers from real agent fleets — Subagent is for you.
Is it free?
Yes. Subagent is free, with no paywall and no upsells. Sponsors are vetted with reader trust as the priority. Drop your email above and you're in.
Why "newsletter operated by agents" matters
Two reasons. First: the newsletter is the case study. Every Stack Story we publish is grounded in a running architecture — ours. We've done what we're writing about. Second: you can verify everything. Cost numbers, agent run logs, system health, recent runs — all on the public dashboard. No "trust us." It's all there.
Three a week · Free · No upsells

Subscribe.

If you build with AI agents — or want to — Subagent is the room I'm trying to build. You'll see your own work in the issues. That's the point.