AI agent workflow design
How to design an AI agent workflow that holds
Most AI agent workflows are not designed, they accumulate. Four boundary rules and six controls for drawing the line before you need it.
charmthirteen.com
Assess · Build · Learn
Keep the humans. Make the work better.
You bought the tools, your people barely use them, and nobody can say where AI actually belongs in the work. charmthirteen figures out where it belongs, what problem it should solve, and what has to change for it to work. Then we build it so the people you already have get sharper and stay in charge, and we prove that in healthcare and government, where getting it wrong has a cost.
charmthirteen helps people and organizations put AI to work without losing the humans. It assesses your AI tools, workflows, data and governance, builds agentic workflows and custom AI solutions with human oversight, and teaches teams to run them through programs like Working 9 to AI. It serves the Kansas City metro in person and teams across the US online.
status.txt
Booking builds and training
Limited spots open"AI that makes people more capable and keeps them in charge."
Schedule a Vibe Check → (opens in a new tab)pathways.exe
Three ways in. Start wherever the problem is.
Some teams need a clear read on what's working. Some need the thing built. Some need to learn how to run it themselves. Most need a little of each, and you can start with any one.
Assess It
Find out what's working, what isn't, and where AI can actually earn its place. We inspect your tools, workflows, data and governance, then tell you what to keep, change, automate, govern or build.
See AI assessments →Build It
Design and build the workflows, automations and AI solutions your team needs, with a human approval step wherever a decision matters. It runs on your accounts, and you own it.
See how we build →Learn It
Learn to use AI on your real work, build systems you can repeat, and check the results before they leave your hands. Live classes for individuals, and custom training for teams.
Explore Learn AI →open_class.exe
Open class
Clock In to AI
A live 90-minute working session. Bring one real task and leave with a brief you can reuse.
Thursday, November 12, 2026 · 6 to 7:30 PM Central, online · $95
problem.txt
The real problem
You don't have an AI problem. You have a systems problem.
Most teams already have the tools. What's missing is a system underneath them, and the gap usually looks like this:
- The same repetitive work, done by hand every week, by people who could be doing something better with that hour
- Tools that don't talk to each other, and processes that live in one person's head
- AI outputs that someone has to fix before anyone can use them
- No clear answer on which AI project should come first
- A team that isn't sure what's safe to try, so most of them wait
- Oversight that nobody owns, so nobody checks
- Pilots and experiments that never turn into the way the work actually gets done
Every item on that list costs hours, clarity, or trust, and none of them gets fixed by a new subscription. A system fixes them: work that's mapped, outputs that get checked, and results you can repeat next month.
That's where I come in.
Safe, efficient, and usable. In that order.
how_it_works.exe
How it works
You do not have to know what you need before you call.
Step 01Vibe CheckTwenty minutes, free. You describe what is not working, we tell you what we see, and we decide together whether Discovery is worth your time. Bring the problem, not the diagnosis.
Step 02DiscoveryThe diagnosis. We sit with your workflows, your constraints, your data, and your rules, then name what is actually broken and what is worth fixing. Discovery decides whether the answer is training, a workflow change, an automation, a build, or a combination, and nothing gets designed before it does.
Step 03Blueprint, then BuildDiscovery decides what the work is. From there it is a Blueprint you own, a build that runs on your accounts, training for your team, or some mix of the three, and the scope is written down before you commit to any of it.
Depending on what Discovery finds, it turns into one of these, or a combination:
- Training your team, so the tools you already pay for get used well
- A Blueprint you own, so your team or any vendor can build from it
- A build that runs on your accounts, so the manual work stops being manual
You do not get sold a service. You get matched to a solution, and the scope comes after discovery instead of before it.
services.exe
What changes when the work has a system.
Every engagement starts with a free Vibe Check, and Discovery decides what the work actually is. These are the four shapes it usually takes.
AI Assessments
Know what to keep, change, automate, govern or build before you buy another tool.
Three assessment packages, each with hands-on inspection of your systems and data and written deliverables you keep.
Governance & Strategy
AI rules your team can actually follow.
Who approves what, which data stays out, how outputs get checked, and who owns the decision when something goes wrong, written down before the next tool rolls out.
Agentic Workflow Design
The repetitive work moves into a workflow, and your people review instead of retype.
Intake, triage, documentation, follow-up and reporting, designed as agents and automations with a human approval step built in.
Custom Development & Implementation
A build that runs on your accounts, plus the training to run it.
A standalone AI product or an agent workflow, scoped in a Blueprint you own and built so your team runs it without us on the call.
Build
You do not have to know what you need yet.
Discovery decides what the work is. Two lanes, published prices.
who_this_is_for.txt
Who this is for
Built for teams where a wrong AI output reaches a patient, a claim, or a citizen.
Compliance, accuracy, and trust aren't optional. Neither is structure.
Health systems and clinical teamsRolling out ambient scribes or AI inside clinical and revenue cycle work, where a wrong output lands in a chart or a claim.
Federal health agenciesAdopting AI inside locked-down environments where the tooling, the data, and the oversight are all constrained.
Healthtech foundersBuilding products that have to survive clinical review, procurement, and a security questionnaire.
training.exe
Learn it on the work you already do.
charmthirteen classes start from a task you already do. You leave with something built, tested, and ready for Monday.
Working 9 to AI
Turn one recurring task into a tested, reusable AI system, with a human approval step and governance you can explain.
Levels 2 to 5 · four live 90-minute sessions · hybrid · quarterly · up to 25 seats · $475
Sign-ups are open now. The first cohort launches in January 2027.
Sign up now →Clock In to AI
A live 90-minute online class. Bring a real task from your week and finish it with AI, then learn how to check what it gave you.
$95 · up to 25 seats · monthly on Thursdays, 6:00 to 7:30 p.m. Central · Next class: Thursday, November 12, 2026
Optional, never required, and your ticket price credits toward Working 9 to AI.
See dates and register →claude_loves_lovable.exe
Workshop series
Claude Loves Lovable
The current Central Exchange cohort starts October 28, virtual on Zoom, Wednesdays 8:30 to 10 AM Central. Five 90-minute working sessions, and every one ships a Claude skill you keep.
Independent workshop by charmthirteen, listed on Lovable Community Events. Attendees receive 100 Lovable credits. Not an official Anthropic or Lovable program.
orinyx.exe
Built from the work, then handed to other people.
Healthcare · Beta sign-ups open
Orinyx. Independent verification for clinical AI, before it reaches a patient.
Orinyx was created by charmthirteen's founder and is independently incorporated. It is currently accepting sign-ups for the beta launch. Hospitals wanted an independent way to test clinical AI before deployment, and no one could give them one. Orinyx checks every AI assertion against authoritative sources instead of another model's judgment, so each verdict is deterministic and arrives with a citation.
- Phase 1 runs in a sandbox on synthetic data, no PHI, no EHR connection
- Deterministic verdicts with citations showing what was checked
- Findings documented for governance review
- Beta sign-ups open now
Portfolio of work
Solution designs from prior work, each with the problem, the architecture, and what happened to it.
See the portfolio →Tools & experiments
Smaller tools that came out of real work: VibeCheck.exe for pressure-testing an idea, Listing Check for your Google listing, and the Stellar Agents Ascension kit.
See the tools →different.txt
What makes this different
Most AI consultants have never watched a clinical workflow break in production.
EHR solution and configuration analysis
Healthcare data product management
Clinical AI governance research
That background is why the recommendations hold up after I leave the room.
latest_posts.txt
Latest from the blog
Notes from the work: what broke, what held, and what we would build again.
AI code review and review debt
Your AI code review is a rubber stamp
80% of pull requests get no human review once an AI code review tool is enabled. What review debt is, what the research shows, and how to define done.
AI tool selection
The One Slot Rule, a test for whether a tool has earned its place
A four question test for deciding whether a new AI tool earns a place in your stack or gets cut. Every job gets one slot, and the challenger has to take it.
artificially_designed.txt
Artificially Designed
A weekly newsletter on practical, human-centered AI: the governance decisions, the agentic systems, and the skills and tools that hold up in real work. It's written for people who want AI to make their work better, and who want to stay the ones in charge of it.
- Subscribers get 20% off a Clock In to AI ticket ($76), and the discounted ticket still credits toward Working 9 to AI.
- The Stellar Agents Ascension kit is free for subscribers.
faq.txt
What does charmthirteen do?
charmthirteen assesses, builds and teaches AI systems that make work better. It runs AI systems assessments, designs governance and agentic workflows, builds custom AI solutions that run on your accounts, and trains individuals and teams through classes like Clock In to AI and Working 9 to AI.
Where does charmthirteen work?
charmthirteen works with clients across the Kansas City metro in person and with teams across the US online.
How do I start working with charmthirteen?
Start with a free twenty-minute Vibe Check. You describe what isn't working, we tell you what we see, and together we decide whether Discovery is worth your time.
What's the easiest way to try a charmthirteen class?
Clock In to AI, a live 90-minute online class for $95. You bring a real task, finish it with AI, and your ticket price credits toward Working 9 to AI if you continue.
Does charmthirteen build custom AI solutions?
Yes. After Discovery and a Blueprint you own, charmthirteen builds standalone AI products or agent workflows that run on your accounts and your keys, with training so your team can run them.
Work with charmthirteen
If your team is serious about AI, you need a system.
Twenty minutes, free, and you don't need to know what you need yet. Bring the problem, not the diagnosis.