Artificially Designed · Issue 22: The Coworker on Another Shift

Issue 22 · The Coworker on Another Shift

If this landed in spam or Promotions, drag it to your inbox and hit reply with one word. It tells your inbox this is mail you want, and it helps every issue after this one reach you.

Hi friend,

Last issue ended with the stack, and I said this one would show you the hardening agent I built for my own workflow. It does, and it starts with a plan I gave up on.

So this week begins with that plan, covers the five places I look before anything built with AI ships, and ends with how I intend to test what Orinyx knows.

This week

🔨 Built Safe, the always-on agent I decided against, five places to look before you ship anything built with AI, and how I plan to test what Orinyx knows

⚡ One Slot, the always-on agent, Devin, Claude Duet and OpenAI's Dots, and my verdict

🌎 In the News, Doximity's dosing research and Claude's read-only finance feature

✅ Before You Ship, the five-places check

💻 Prompt for Productivity, audit your code for the five places

🗓 Open Calls, Claude Loves Lovable, the Orinyx MCP through 0.1.4, and four picks, a Kansas City data science night, HLTH in Las Vegas, a Kansas City AI meetup and Stanford's virtual AI+HEALTH

🔭 Next From Me, knowledge graphs, ontologies and CHARM

I planned to build my agents as always-on agents on Claude's SDK, workers that sit in their own cloud environment and wait for something to do. I had the fleet designed and the roles named, and then I priced it. An always-on fleet spends tokens whether or not anything useful is happening, and I could see how quickly I would be paying for agents to wait. I like how my fleet and my sub agents build, so I kept that design and changed where the work happens.

Devin covers the part I wanted. It works in its own cloud environment, you can configure it to run GitHub Actions, and it keeps going on a ticket while you are away from your screen. My spot on the Devin iOS beta came through this weekend, which means I can check on it and steer it from my phone. I am the maintainer of the repo and I still have to be away from my desk, so I wanted a person in the loop on every change, and that became one rule.

The one-shot rule. The builder gets one pass at a ticket, and I review the result before anything moves on.

Here is how the work is split now.

Claude plans. I use it for planning, design, ideating and brainstorming, and for deciding what to delegate and to whom.

Devin builds. It works the ticket in its own environment.

Codex reviews. It reads the code Devin wrote, so a different tool checks the first one.

I decide. I read it, clean it up and push.

The review step looks like an LLM as judge with me in the middle, and I want to be exact about it. Codex judges code. It does not judge whether a medication claim is right, because in Orinyx the verdicts come from deterministic rules and cited FDA label text, with no model deciding. Last issue I wrote that the system that builds a checker should not be the only authority grading it, and this is the same rule applied to my own code.

None of this needs agents. If you build with Lovable, Claude Code, Codex or Devin, or you paste code out of a chat window, you have the same gap, because generated code looks finished before it is safe. I look in five places every time, and I do not let the sweep improvise.

Each finding gets a file, a line and a way it could be attacked, and a finding without those is a note and not a finding. Each fix gets a test that fails if the hole ever reopens, because a fix without that test is a hope. And whoever fixes it should not be the one who says it passed.

What I am designing now. I am mapping Orinyx's medication batches, and batches 2 and 4 are in flight, the ISMP chemotherapy IV drugs and the Top 300 cardiometabolic drugs, both due November 20. After the mapping comes the data schema that has to hold each batch as it releases, and after each release come agents that test it. Those agents work from a set of questions that generates test questions for me to check, and I plan to model the architecture on HealthBench, OpenAI's open health benchmark on Hugging Face, where each answer is graded against physician-written rubric criteria. The point is to confirm that the knowledge living inside Orinyx is accurate, and not only that the checker runs. HealthBench grades with a model, and in Orinyx no model sits in a verdict path, so these questions test Orinyx and never decide a verdict.

Three moves came out of this, and none of them need Orinyx.

Give the builder one pass. Review it before anything moves on, so a person is in the loop on every change.

Look in the same five places every time. The same sweep catches the same kinds of holes.

Make every fix earn a test. If the hole reopens, the test fails.

The One Slot Rule is that every job gets one slot, and a new tool has to take it from whatever holds it now. This week the slot is the always-on agent, and three tools are asking for it.

What is the job. Working on something while I am not at the screen, in an environment of its own. Devin holds the build version of that slot for me today. Anthropic's Claude Duet launched on October 21 as a proactive agent that connects to your apps and can work in the cloud in a virtual machine with its own browser, and OpenAI's Dots, announced at DevDay on September 29, are always-on agents that work in the background on their own cloud computers.

Can it do that job, not can it do more. Devin can, for building. Duet is a research preview that opened by waitlist to Pro and Max subscribers in the United States, and it is still in beta for me. Dots started rolling out to ChatGPT Pro subscribers. Both are early, and I am trying both before I say more than that.

What does it touch that is hard to undo. For Duet it is the list Anthropic itself names, sending a message, making a payment, signing a document and deleting a file, and it asks for approval before each one. Dots ask for approval before sensitive actions like installing software or changing a password, and they have a Custom Rules tool for boundaries you set. For Devin it is a change to my repo, which is why the one-shot rule exists.

What would have to be true to let it run unwatched. An approval gate on every action that cannot be undone, and a person reviewing before anything merges. Both Duet and Dots already build the first half in, and the second half is a habit I have to keep.

My verdict is Devin. It holds the slot, because it is the one of the three that already does the job I need, and Duet and Dots stay on trial until one of them earns a job of its own.

Two stories from the last two weeks, and both are about how much access an AI gets and who checks it.

💊 AI models get medication doses wrong, and a vendor says it has a check for that. On October 29, Doximity published research in which nine widely used AI models answered more than 8,500 clinical dosing questions covering more than 2,200 FDA-approved drugs. About 19.5 answers per 1,000 contained an error that was serious or potentially lethal, and the most common mistake was failing to adjust a dose for age, kidney function or drug interactions even when the question included it. Doximity says its new DoseCheck tool got dosing findings right 99.9% of the time. These are the company's own numbers for its own product, and I would want an independent evaluation before leaning on them, which is also the standard I hold my own work to. A rule-based check against FDA labeling is the same bet I am making with Orinyx, so I will be reading their methods closely. Healthcare Dive's report

🏦 Claude can now read your bank accounts, and only read them. On November 5, Anthropic added Personal Finance to Claude, which lets you connect bank and card accounts for analysis. Access is read-only, and there is no option for Claude to move, manage or invest money. I like the order of it, trust gets earned in steps and the irreversible actions stay with the person. MacRumors' report

Run the five-places check on the last thing you built with AI. Open the code and answer each question with a file and a line, whether it is secrets, access, input, boundaries or code rot. If an answer starts with "I think so," that question fails.

If you can point at the line, you have an answer. If you need to ask the tool whether it did it, you have a dependency, and the smallest fix is usually to ask the tool to show you the line and then to write the test that proves it.

Audit your code for the five places.

WHEN TO USE IT. Before you share, deploy or hand off anything an AI tool built for you, and again after any large change.

WHAT IT DOES. It checks the same five places every time and makes every finding point at a file and a line, so you can verify it instead of trusting it.

TIP. Replace every real key with a placeholder before you paste, because a security audit should never be the thing that leaks the secret. Then take each Part 1 finding and ask for a test that fails if the hole reopens.

From me

Wednesday, November 11, 8:30 to 10am Central. Claude Loves Lovable, my hands-on workshop series with Central Exchange, where we build with Claude and Lovable, with session three building an MVP. Sessions run November 11 and 18 plus an optional one on December 2, and registering for one covers all of them, so please register only if you plan to attend live. It is virtual, no coding or prior Claude experience is needed, and it is free for Central Exchange members and $25 for non-members. Register through Central Exchange

The Orinyx MCP, now through 0.1.4, and beta sign-ups are open. It checks the medication orders and claims your agent produces against FDA label text, cites the section it checked, and says plainly when it could not check. Since last month every result row names the check that produced it, developer accounts can use five FDA label and monograph lookups that do not count against the free plan's 200 monthly checks, and there is now a troubleshooting guide and a stdio command for setup in Claude Code, Claude Desktop and Cursor. It is a limited testing release on synthetic data, not for clinical use. Join the Orinyx MCP waitlist

Four picks for the week

Thursday, November 12, 6pm Central. Data Science KC University & Early Career Showcase, free and in person at Burns & McDonnell, 9450 Ward Parkway in Kansas City, with lightning talks from regional students and new graduates, including one on evaluating clinical trial consent forms with LLMs. Details

November 15 to 18. HLTH USA 2026, in person at The Venetian Expo in Las Vegas, a large healthcare innovation event. See the details

Thursday, November 19, 6 to 8pm Central. Global AI Kansas City, the chapter's November meetup, in person in Kansas City. Event page

December 8 and 9, 8am to 2:30pm Pacific. Stanford AI+HEALTH 2026, a virtual conference on AI in health and medicine, $350 for academic, healthcare, nonprofit and government attendees, $550 for corporate, and $50 for students, with limited complimentary scholarships for people with financial need whose applications are due November 15. Register

Tool verdicts, templates and reference material live on the site at charmthirteen.com.

Next week I go back to knowledge graphs and ontologies, what they are for and why I keep reaching for them, and I owe you the proper layout of CHARM, which I will fold in.

What is the one thing you would never let an AI do without asking you first? Hit reply and tell me, because the answers shape where I put the next approval gate.

Lex