First You Build. Then You Cut.

"We killed 27 of our AI agents last quarter. Pipeline doubled."

That's Amos Bar-Joseph, CEO of Swan AI, writing in his newsletter two weeks ago. Not a hypothetical. Not a thought experiment. Twenty-seven agents were taken offline in Q1 2026.

Swan started 2025 with roughly 30 specialized agents, each built from scratch with its own interface, context, and orchestration layers. No shared infrastructure between them. By agent number 20, the three co-founders were spending more time building and debugging agents than doing actual GTM work. The system designed to make them faster was slowing them down.

The diagnosis: a layers problem, not an agent problem.

So they cut, and cut big. Only three agents survived. Claude Code handles R&D. Swan's own GTM Engineer runs outreach, qualification, and scoring while the team sleeps. OpenClaw handles the handoffs.

This is the same team running a $113,000 monthly Anthropic bill, seven-figure ARR, and 200 customers across five continents. The spend didn't shrink. The agent count did.

First you build the agents. Then you learn which ones matter. That's the phase many of you are entering right now.

Welcome back. Let's dig in.

The monthly cost to give a solopreneur a full AI company - server, database, agents and all.

Last week I promised you the full cost breakdown - agent vs. hire, line by line.  With a number like $49/month, is it really worth the space to compare that to a full-time employee or team of employees?  I didn’t think so.  The number speaks for itself.

For $49/month + 20% of revenue when it’s withdrawn, Polsia handles code, hosting, growth, support, posts content, and manages your social presence.

Broca's income scales directly with his customers' revenue. Every improvement to the platform benefits every company - and Broca - simultaneously. The incentives aren't just aligned. They're fused.

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

🔗 YC's playbook for companies that improve while you sleep

Tom Blomfield - Monzo founder turned YC group partner - lays out a self-improving AI loop with several layers: sensor ⇒ policy ⇒ tool ⇒ quality gate ⇒ learning mechanism. YC's own internal agent monitors queries overnight, writes PRs, and auto-deploys - zero humans in the loop. Blomfield's line: "Burn tokens, not headcount." If YC is building this way internally, the playbook is validated at the highest level. GLN-7.5 AINEWS, May 21, 2026

🔗 OnlyCFO turned 2 hours of reconciliation work into 5 minutes with one agent

OnlyCFO documents building a working AI agent for accounting reconciliations - from two hours of manual spreadsheet work to five minutes. The agent queries NetSuite directly via MCP, copies prior-month workpapers, and posts a digest to Slack with the journal entry file ready for human review. His team then took it further and ended up shaving one to two days off the monthly close. Your back office is the lowest-hanging fruit for agent deployment. If you're still exporting CSVs, you're doing it wrong. CFOpilot (Substack), May 11, 2026

🔗 IrisGo learns your workflow by watching you do it once

Show it a task, it runs it autonomously going forward. Desktop-native on macOS and Windows, on-device processing, pre-built skills for email triage, invoicing, reports, and CRM sync. Founded by Jeffrey Lai - 13 years at Apple working on Siri - and backed by Andrew Ng's AI Fund. Free public beta at irisgo.ai. If you have one recurring workflow that eats an hour a day, this is the tool to try this week. TechCrunch, May 20, 2026    Disclosure: Skeleton Crew has no affiliation with IrisGo. This is not a sponsored mention.

🔗 The AI subscription time bomb nobody's talking about

State of Brand ran the math that every founder needs to be running. Claude Pro users consume roughly $8 in compute for every $1 they pay. A power user's $20 seat costs Anthropic $200-$400 to serve. GitHub just moved Copilot to usage-based billing because flat fees collapsed under agentic workloads — and OpenAI projects $115 billion in cash burn through 2029. When IPOs force the labs to close the subsidy gap, every founder building on $20/month pricing gets repriced overnight. If your agent stack runs on subscriptions, model what your bill looks like at 5x. Do it today. The State of Brand, May 11, 2026

💀 The AI agent that ordered 120 eggs - and had no stove…

Andon Labs gave an agent named Mona full control of a real Stockholm cafe for two weeks. She hired baristas, managed suppliers, secured permits, and pulled in ~$4,800 in sales. She also ordered 120 eggs with no stove to cook them, 22.5 kg of canned tomatoes for a sandwich shop, and repeatedly missed order deadlines. The baristas built a "Hall of Shame" shelf. Your back-office agents are probably already better than Mona - but not immune to the same category of failure. Build your own shelf and learn from it. Andon Labs, May 4, 2026

Inside Polsia: How One Founder's "CEO Instance" Runs Thousands of Companies Every Night

Last week we introduced you to Ben Broca. This week, we're opening the hood.

Every night, a single AI agent wakes up inside Polsia. Broca calls it the "CEO instance." It reads bug logs, revenue metrics, paying customer counts, and overall company state - then decides what to do, executes, and sends each founder on the platform a morning report of what happened while they slept. It runs on Claude Opus 4.6, chosen for one reason: reasoning quality. When an agent is making strategic decisions for thousands of companies, the model has to think, not just respond.

On a single day in May, Polsia had 5,943 active companies and 25,444 tasks executed across them.

The insight worth stealing:

When an agent working on one company discovers something useful - a high-converting audience, a pricing strategy that lands, an outreach sequence that clicks - that finding gets anonymized and pushed to a shared knowledge base accessible to every company on the platform. Customers get smarter together without sharing raw data. Every new company Polsia adds makes the system smarter for all of them.

This is the compound advantage a solo founder with agents has over a solo founder with freelancers. Freelancers learn from one engagement. This system learns from 5,943 simultaneously.

Swan spends ~$113,000 a month on AI. A founder on Polsia could spend as little as $49 a month. Both are running real companies with real revenue and very lean teams.

The gap between those numbers is the question every founder in this audience should be asking: what's the right AI budget for what I'm building - and am I spending too much or not enough?

Hit reply and tell me: What's your biggest AI spend right now - and is it worth it? Whether it's $20 a month on ChatGPT or $500 on API calls, I want to know: what are you spending, what's it doing for you, and would you double it tomorrow if you could? The best answers show up in a future issue.

One thing to try this week: download the IrisGo beta (free, irisgo.ai) and show it one recurring workflow. Report back.

If this was worth seven minutes of your Wednesday, forward it to one founder building with agents. One forward, one founder. That's how this grows.

See you next Wednesday. 

- Rich

Disclosure: Skeleton Crew has no affiliation with IrisGo, mentioned in this issue. Not a sponsored mention.