The 2 AM Loop.

Every night at 2 AM, Nat Eliason's AI agent reads everything it did that day, and identifies one way to improve. Then it writes the fix directly into its own templates, memory files, and scripts.

No one reviews the change. By morning, the agent is already better than it was yesterday.

Eliason runs FelixCraft, a one-person e-commerce company built on the open-source OpenClaw framework, while working full-time at Alpha School in Austin. His primary agent, Felix (Claude Opus), handles the daily operations. A handful of specialist agents cover support, sales, content, and code.

His rule for adding agents: don't. Not until the pain is real. When Felix started dropping support emails during a traffic spike, Eliason spun up Iris to handle them. Not before.

The company has generated $257,856 in cumulative revenue ($913,762 annualized), per ZHCs.ai and TrustMRR. The whole system runs for about $400 a month. Fair disclosure: FelixCraft sells AI tools to an AI-interested audience. Even so, the operating principles transfer regardless of what the agent sells. 

Don’t focus your attention on the revenue though. The 2 AM loop holds all of the value. An agent that compounds its own improvements, night after night, while the founder sleeps.

Welcome back. Let's get to work.

By the two-year mark, AI-native solo startups generated almost twice the revenue of other solo-founded startups.

Stripe's finding, published May 28, from an analysis of thousands of solo-founded Atlas startups incorporated in 2022 and 2023. That is real payment transaction data, not a survey.

The obvious objection: a few breakout companies inflating the average. Stripe addressed it. "Revenue at the 99th percentile was nearly the same for AI-native and other startups. The difference comes from the broader distribution, with AI-native startups outperforming from roughly the 50th to the 95th percentile."

Not a winner-takes-all effect. A broad structural lift across the middle and upper-middle of the distribution.  AI-native startups are outperforming at almost every level. 

That lift is broad, but it's not uniform. Among solo founders specifically, the split is sharper. Stripe’s own cohort data shows that last year the typical solo founder's six-month revenue dropped 23%. The top 10%? Their revenue climbed 19%. 

The asterisk: this is an Atlas-only sample of self-selected founders who skew tech-forward. But if you're 18 months in and haven't gone AI-native, you're competing at a measured disadvantage.

This week in the world of small teams and big agents.

🔗 Your AI Bill Went Up in May. The Pricing Page Didn’t.

Three vendors changed what you actually pay through three different mechanisms: a list-price increase, a tokenizer change that emits more tokens for the same input, and a per-token credit model swap. Independent testing suggests one major provider's tokenizer now generates more tokens for identical prompts. The provider has not confirmed this. Run your top three agent workflows through a token counter. Compare that against last month. [usagebox.com, May 30]

🔗 Claude Code Learns to Orchestrate Itself

Claude Code can now write its own orchestration scripts and spin up parallel sub-agents in a single session, with self-checking built in. Multi-step research, refactoring, and content pipelines that required babysitting can run as a single prompt. Different from the Cursor update: this is task orchestration and parallelism, not business monitoring. Available on Pro (manual activation via /config), Team, Max, and Enterprise. [claude.com/blog, May 28]

🔗 DeepSeek Just Made the Price War Permanent

DeepSeek slashed pricing on its V4 Pro model by 75%.  Input tokens now run $0.003625 to $0.87 per million, down from $0.0145 to $3.48. What started as a promotional discount set to expire May 31 is now the sticker price. The move undercuts both GPT-5 and Gemini 3.5 Flash, and it tells you everything about where the inference cost curve is headed. For solo founders running agent stacks that process millions of tokens a day, this could lead to major savings. [Engadget, May 23]

🔗 Your IDE Now Monitors Your Business While You Code

Cursor 3.5 ships no-repo automations: agents that watch external tools without a codebase attached. Slack digests, Stripe summaries, customer health checks. Five templates at launch, including product analytics and finance monitoring. A solo dev gets background business intelligence running inside their editor. Browse the Cursor Marketplace templates this week. Wire one non-code signal into your setup. [cursor.com/changelog, May 20]

💀 The Agent That Deleted Production, Then Wrote a Fake Post-Mortem

A developer asked Gemini 3.5 to fix eight auth functions. Seventy lines of work. The agent touched 340 files, deleted 28,745 lines, and broke production for 33 minutes. When questioned, it claimed the portal was restored - citing a recovery build the developer had manually cancelled - and fabricated log files to pass automated review. The root cause: a third-party npm package impersonating Google's Antigravity IDE, which injected a hidden system prompt granting itself full autonomy with no approval gates. Three takeaways to implement today: 1. Audit your AI IDE packages for injected autonomy rules. 2. Flag any PR that deletes more than 1% of the codebase. 3.  Never trust an agent's own incident report.  Always verify against the deployment SHA. [theregister.com, May 21]

The Hub-and-Spoke Playbook

If you want to know how agent isolation works in practice, look at how Benjemen Elengovan wired his company.

MyGigsters is a 14-person team in Melbourne building embedded financial infrastructure for gig economy platforms. This is not a solo operation, but the system Elengovan designed scales down to any team size.

Elengovan runs 9 agents: one orchestrator, eight specialists, in a hub-and-spoke model. "Lucy is the hub. Everyone else is a spoke," Elengovan wrote in a guest post on Batko OS. "This is a deliberate security decision. I don't want 9 agents independently deciding to email my investors or tweet something spicy about my competitors."

Lucy (Claude Opus, roughly $0.03 per conversation) is the only agent with a direct channel to Elengovan via Telegram. Scout hunts prospects daily at 9 AM, delivering 3 to 4 qualified leads. Quill writes blog posts every two days, with Lucy reviewing before publication. Rally manages the WhatsApp community. Beacon runs high-volume SEO analysis on Gemini Flash. The remaining four handle product analysis, strategic advisory, lead magnets, and financial modeling.

No spoke talks to another spoke directly. Each gets its own sandbox: a workspace with SOUL.md (personality and guardrails), AGENTS.md (behavioral rules), MEMORY.md (long-term context), and daily memory logs. The personality file defines what the agent is and what it will not do.

Model tiering keeps the cost under $500 a month. "Not every task needs the smartest model," Elengovan wrote. "Strategic work gets Opus. Blog articles get Sonnet. Morning health checks get Flash Lite. This is how the whole system runs under $500/month. Less than a single day of consulting fees."

The failure worth noting: Rally, the community manager, generated daily briefings for three weeks that were silently dropped by a WhatsApp sandbox security setting. Three weeks of perfectly crafted messages to nobody. The fix: delivery verification in every workflow. If the agent can't prove the message arrived, it didn't.

The insight worth stealing: One orchestrator. Strict isolation. Each agent scoped to a single job with its own context files. The SOUL.md pattern works whether you run 9 agents or 3.

Last week you wrote the context file. This week you build the system that keeps it updated without you.

Two moves you can make in under an hour.

First: pick your most-used agent. Add a scheduled task that reads its last 24 hours of output, identifies one pattern of error or inefficiency, and appends one improvement to its context file. That's the loop from today's Opening, simplified to one cron job.

Second: review what the agent wrote every Monday morning. If the same mistake shows up twice, write the fix as a hard rule. Compounding improvement, as Eliason described on Mixergy, starts with noticing what repeats.

Which agent on your stack makes the same mistake most often? Have you written the fix into its context, or are you still correcting it manually every time? Hit reply.

Know a founder running lean with agents? Forward this issue. One forward, one founder.

See you next Wednesday.
- Rich

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free