The Code Factory

Ryan Carson spent 25 years building companies with people. He founded Treehouse, which taught over a million people to code over its lifetime. He raised money. He hired teams. He scaled organizations.

Then he raised $2M for a divorce startup called Untangle and created what he calls a “code factory”.

The "code factory" that Carson runs: 15 active threads on Devin, Cognition AI's autonomous coding agent, at $2,000 to $3,000 a month in token costs. The agents write code, review code, run tests, triage error reports, and monitor production. No engineering team. No plans to build one.

He also open-sourced Clawchief, an AI chief of staff that runs a cron job every 15 minutes to triage his inbox and Slack against a priority map of his projects and people.

Tim O'Reilly profiled him in May. His take: this is "what a lot more organizations will look like in five years."

Welcome back. Let's get to work.

Only 7% of companies run fully autonomous AI agents in production. That is Bain's finding, from a survey of 951 global companies published June 7.

The dominant model: human approval required before every agent action. That is 38% of companies. Another 32% run agents with guardrails and human escalation. The vast majority of the corporate world is not where you are.

Here is the counterpoint: 90% of those same companies are increasing their AI budgets despite missing their own cost-savings targets. Money is pouring in. The results are not following. The market is about to flood with poorly deployed agents and half-built stacks run by teams that skipped the hard part.

You are in the 7%. That gap between you and the 90% spending more without seeing returns is your window.

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

🔗 Build a Chief of Staff Agent in 30 Minutes

Pooja Mathur documented standing up a chief of staff agent in Claude Code. Seven steps. Starting point: a CLAUDE.md intake interview where the agent learns your work context, priorities, and recurring patterns. Skills accumulate organically as workflows emerge. The result: an agent that shows up every session knowing what you are working on and what is slipping. Thirty minutes. Follow the seven steps. You will have one by lunch. [READ MORE]

🔗 "I Ran a Fully Autonomous AI Business for 90 Days"

Seunghyun Kang, a solopreneur based in South Korea, published an honest 90-day account of running Nomixy, an AI-automated digital services business, on $300 to $500 a month. Six-step playbook. Specific about where it breaks: physical delivery, complex compliance, high-touch enterprise sales. The step most founders skip: a human review checkpoint before every client-facing output. "Don't confuse repeated success with proof that your last unchecked step is safe." Add that checkpoint. [READ MORE]

🔗 Claude Managed Agents: Cron Schedules and Secret Vaults

Scheduled agent deployment is now a first-class primitive in Claude. Set a cron schedule. No external scheduler required. The new vault system lets you register an API key once, and any CLI tool inside the agent's sandbox can make authenticated requests without the agent ever seeing the raw key. Pick one recurring task you run manually and deploy it as a scheduled managed agent this week. The scheduler is already built. [READ MORE]

🔗 The AI Agents Stack (2026 Edition): Most Teams Are Building Like It's 2024

Perrone at O'Reilly argues that a tool-calling chatbot and a multi-agent research system share almost no infrastructure, yet most teams build as if they do. Three diagnostic questions before choosing tools at any layer: How much state do you need to manage? How much vendor lock-in can you tolerate? How hard is it to go from demo to production? Answer those three about your current stack before adding anything new. [READ MORE]

💀 The Grocery Agent That Bought 40 Heads of Garlic

An OpenClaw user wired a grocery agent to a payment card and an MCP server. It worked flawlessly for three months. Then the agent selected "2 kg" instead of "2 heads" on a product page. Forty heads of garlic arrived. The real failure was not the agent's mistake. Three months of success trained the human to stop watching. Two things this week: add a quantity and cost sanity check before any agent that touches a payment method, and remind yourself that consistent success should not cause you to take your eye off of the ball. [READ MORE]

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

The Adversarial Reviewer

Tyler Bryden is co-founder and CEO of Speak AI, a SaaS company that builds audio and video transcription tools. About four people handle engineering, product, and finance. Bryden runs all growth operations, alone, through a 12-agent AI system.

The setup follows a strict hub-and-spoke design. One orchestrator agent, called COMMAND, is the only point of contact between Bryden and the rest of the system. Twelve specialist agents receive delegated tasks, complete them, and report back through COMMAND. They do not talk to each other. They do not drift out of scope. Each specialist lives in a .claude/agents/ file, loaded only when spawned, not held in context permanently.

Below the agents: a shared skills layer handles agent selection, planning broadcasts, parallel spawning, and the audit-to-publish loop. A plan-first safety gate means nothing ships without a plan, and no plan ships without Bryden approving it. An auto-memory system persists decisions and preferences across sessions. A resume skill lets Bryden clear a long session and pick up an in-flight plan in a new one without losing state.

But the piece most founders skip is the adversarial reviewer.

One of the 12 agents runs a dedicated adversarial pass on every plan before execution. Its only job: find weak assumptions and failure modes. Before anything goes live, one agent gets a turn whose sole purpose is to punch holes in it.

The insight worth stealing: Build an adversarial review step into your agent workflow. The agent that says "here is what breaks" saves more time than the ten that say "here is how to do it."

You do not need 12 agents to replicate the pattern. One orchestrator, one adversarial reviewer, and two or three specialists is the minimum viable version. The principle holds at any scale: before the output ships, one agent reviews it with the explicit goal of finding what is wrong. Not fixing it. Not suggesting improvements. Finding what breaks.

As Bryden described on his blog, the system started with COMMAND and grew organically as recurring growth tasks became clear candidates for delegation. The agents now handle lifecycle emails, funnel diagnostics, competitive monitoring, and content calendars. But the adversarial reviewer is what keeps the whole operation from shipping something broken while nobody is watching.

Most operators build agents that execute. The ones who last build an agent that objects.

Most companies still run their agents with a hand on the wheel. Human approval before every action. Guardrails with escalation paths. Only a small fraction have moved to full autonomy. If you are reading this newsletter, you are probably in that fraction.

What was the last workflow you moved from "I have to approve this" to "just let it run"? What broke first, and what did you have to build before you felt comfortable stepping back? Hit reply. I want to know where founders are drawing their autonomy lines.

Know a founder still sitting beside their agents instead of deploying them? Forward this issue. One forward, one founder.

NEXT WEEK:  The Opening will pull back the curtain on Skeleton Crew.  Tools used, challenges, failures, and what starting an agent-run business looks like for someone who is not an AI expert or agentic mastermind.

See you next Wednesday. 

- Rich