https://www.facebook.com/2223178651032431/

Jothi Moorthy AI Advisor, 777 Woodward Avenue, Suite 800, Detroit, MI (2026)

10/10/2026

𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 𝗶𝗻𝘃𝗶𝘁𝗲𝗱 𝗺𝗲 𝘁𝗼 𝗮𝗱𝗱 "𝗚𝗲𝘁 𝗮𝗱𝘃𝗶𝗰𝗲 𝗳𝗿𝗼𝗺 𝗺𝗲" 𝘁𝗼 𝗺𝘆 𝗽𝗿𝗼𝗳𝗶𝗹𝗲.
The invite went to "a small group of trusted industry leaders."
⠀
LinkedIn calls it "a profile button that lets you host paid 1-on-1 meetings."
𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 𝗶𝘀 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗮 𝗻𝗲𝘄 𝗽𝗮𝗴𝗲 𝘁𝗼 𝗳𝗲𝗮𝘁𝘂𝗿𝗲 𝘁𝗵𝗲 𝗲𝘅𝗽𝗲𝗿𝘁𝘀 𝘄𝗵𝗼 𝗼𝗳𝗳𝗲𝗿 𝘁𝗵𝗲𝗺.
The feature is still in beta.
⠀
𝗜 𝘀𝗮𝗶𝗱 𝘆𝗲𝘀, 𝗮𝗻𝗱 𝘁𝗵𝗲 𝗯𝘂𝘁𝘁𝗼𝗻 𝗶𝘀 𝗻𝗼𝘄 𝗼𝗻 𝗺𝘆 𝗽𝗿𝗼𝗳𝗶𝗹𝗲.
Thank you, LinkedIn team, for the invitation.
⠀
𝗪𝗮𝗻𝘁 𝘁𝗼 𝘀𝘁𝗮𝗻𝗱 𝗼𝘂𝘁 𝗮𝘀 𝗮 𝘁𝗵𝗼𝘂𝗴𝗵𝘁 𝗹𝗲𝗮𝗱𝗲𝗿 𝗶𝗻 𝘆𝗼𝘂𝗿 𝗳𝗶𝗲𝗹𝗱?
Book 30 minutes with me and bring one question.
Your headline, your About section, your next post.
𝗬𝗼𝘂 𝗹𝗲𝗮𝘃𝗲 𝘄𝗶𝘁𝗵 𝗮 𝘀𝘁𝗿𝗮𝗶𝗴𝗵𝘁 𝗮𝗻𝘀𝘄𝗲𝗿 𝗮𝗻𝗱 𝘆𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝘀𝘁𝗲𝗽.
⠀
LinkedIn currently limits hosting and booking to US-based members.
→ The button is on my profile: https://www.linkedin.com/in/jothi-moorthy
⠀
𝗪𝗵𝗮𝘁 𝘄𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗮𝘀𝗸 𝗮𝗯𝗼𝘂𝘁 𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 𝗼𝘂𝘁 𝗶𝗻 𝘆𝗼𝘂𝗿 𝗳𝗶𝗲𝗹𝗱?

10/09/2026

𝗙𝗥𝗘𝗘: 𝟮𝟬𝟬+ 𝗰𝗼𝘂𝗿𝘀𝗲𝘀 𝗼𝗻 𝗩𝗲𝗿𝗶𝘇𝗼𝗻'𝘀 𝗔𝗜 𝗦𝗸𝗶𝗹𝗹𝘀 𝗳𝗼𝗿 𝗔𝗺𝗲𝗿𝗶𝗰𝗮, 𝗳𝗿𝗼𝗺 𝗔𝗜 𝗯𝗮𝘀𝗶𝗰𝘀 𝘁𝗼 𝗰𝗮𝗿𝗲𝗲𝗿-𝗳𝗼𝗰𝘂𝘀𝗲𝗱 𝘁𝗿𝗮𝗰𝗸𝘀.
No Verizon commitment, open to US adults 18+.
⠀
𝗩𝗲𝗿𝗶𝘇𝗼𝗻 𝘀𝗮𝘆𝘀 𝗽𝗿𝗲𝗺𝗶𝘂𝗺 𝗔𝗜 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗰𝗮𝗻 𝘁𝗼𝗽 $𝟳𝟬𝟬 𝗮 𝗽𝗲𝗿𝘀𝗼𝗻 𝗮 𝘆𝗲𝗮𝗿.
𝗧𝗵𝗶𝘀 𝗼𝗻𝗲'𝘀 𝗹𝗲𝗮𝗿𝗻𝗲𝗿 𝗽𝗮𝗴𝗲 𝘀𝗮𝘆𝘀 𝟭𝟬𝟬% 𝗳𝗿𝗲𝗲, 𝗮𝗻𝗱 𝗮𝘀𝗸𝘀 𝗼𝗻𝗹𝘆 𝗳𝗼𝗿 𝗮 𝗳𝗿𝗲𝗲 𝗮𝗰𝗰𝗼𝘂𝗻𝘁.
⠀
𝗧𝗵𝗲 𝗵𝘂𝗯 𝗶𝘀 𝗮 𝗳𝗿𝗼𝗻𝘁 𝗱𝗼𝗼𝗿 𝘄𝗶𝘁𝗵 𝘁𝗵𝗿𝗲𝗲 𝗽𝗮𝘁𝗵𝘀, 𝗻𝗼𝘁 𝗼𝗻𝗲 𝗰𝗮𝘁𝗮𝗹𝗼𝗴𝘂𝗲.
→ One for learners, from job seekers to displaced workers.
→ One for small business owners.
→ And one for K-12 educators.
⠀
𝗧𝗵𝗲 𝗰𝗼𝘂𝗿𝘀𝗲𝘀 𝗿𝘂𝗻 𝗳𝗿𝗼𝗺 𝗼𝗻𝗲-𝗵𝗼𝘂𝗿 𝘀𝘁𝗮𝗿𝘁𝗲𝗿𝘀 𝘁𝗼 𝗳𝘂𝗹𝗹 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻𝘀.
→ AI for Brainstorming & Planning, from Google, one hour.
→ AI for Productivity, from Duke University, five hours.
→ Google Workspace with Gemini Specialization, from Google Cloud, 12+ hours.
⠀
𝗩𝗲𝗿𝗶𝘇𝗼𝗻 𝗮𝗻𝗻𝗼𝘂𝗻𝗰𝗲𝗱 𝗶𝘁 𝗼𝗻 𝗦𝗲𝗽𝘁𝗲𝗺𝗯𝗲𝗿 𝟮𝟯, 𝘄𝗶𝘁𝗵 $𝟳𝟬𝗠 𝗯𝗲𝗵𝗶𝗻𝗱 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝗶𝗻𝗶𝘁𝗶𝗮𝘁𝗶𝘃𝗲.
That money funds the training plus in-person coaching in select regional markets.
Verizon says partners include IBM, Google, Microsoft, Anthropic, Coursera, and OpenAI.
⠀
𝗜𝗳 𝗽𝗿𝗶𝗰𝗲 𝘄𝗮𝘀 𝘁𝗵𝗲 𝗿𝗲𝗮𝘀𝗼𝗻 𝗔𝗜 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗸𝗲𝗽𝘁 𝘄𝗮𝗶𝘁𝗶𝗻𝗴, 𝘁𝗵𝗮𝘁 𝗿𝗲𝗮𝘀𝗼𝗻 𝗶𝘀 𝗴𝗼𝗻𝗲.
⠀
→ Start at the hub here: https://www.verizon.com/aiskillsforamerica
→ Save this for your learning list.
⠀
𝗪𝗵𝗶𝗰𝗵 𝗽𝗮𝘁𝗵 𝘄𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗼𝗽𝗲𝗻 𝗳𝗶𝗿𝘀𝘁: 𝗹𝗲𝗮𝗿𝗻𝗲𝗿, 𝘀𝗺𝗮𝗹𝗹 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀, 𝗼𝗿 𝗲𝗱𝘂𝗰𝗮𝘁𝗼𝗿?

10/08/2026

𝗔 𝘁𝗼𝗸𝗲𝗻 𝗽𝗿𝗶𝗰𝗲 𝗶𝘀 𝗻𝗼𝘁 𝘁𝗵𝗲 𝗰𝗼𝘀𝘁 𝗼𝗳 𝘁𝗵𝗲 𝘄𝗼𝗿𝗸.
Anthropic's own developers now publish free playbooks on claude.dev.
Two of them matter if you own the budget.
⠀
The pricing playbook proves it.
𝗧𝘄𝗼 𝗺𝗼𝗱𝗲𝗹𝘀 𝗮𝘁 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝘁𝗼𝗸𝗲𝗻 𝗽𝗿𝗶𝗰𝗲 𝗰𝗮𝗻 𝗰𝗼𝘀𝘁 𝘃𝗲𝗿𝘆 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗮𝗺𝗼𝘂𝗻𝘁𝘀 𝗼𝗻 𝗼𝗻𝗲 𝘁𝗮𝘀𝗸.
One model does the job in one pass.
The other reads, tries a fix, and reads again.
Each pass resends everything before it, so more passes means a bigger bill.
⠀
Run one real task from your backlog on two models and compare the two bills.
"𝗬𝗼𝘂𝗿 𝗼𝘄𝗻 𝗻𝘂𝗺𝗯𝗲𝗿𝘀 𝗮𝗿𝗲 𝘁𝗵𝗲 𝗼𝗻𝗲𝘀 𝘁𝗼 𝘁𝗿𝘂𝘀𝘁," 𝘀𝗮𝘆𝘀 𝘁𝗵𝗲 𝗽𝗹𝗮𝘆𝗯𝗼𝗼𝗸.
It also reports what companies actually spend on Claude Code.
The average: about $13 per developer per active day, across enterprise deployments.
For 90% of users, it stays under $30 per active day.
⠀
The second playbook puts a number on whether an AI change helped.
In its worked example, a customer support AI was improved.
Then it was scored on 14 tickets the improvement never touched.
𝗔𝗰𝗰𝘂𝗿𝗮𝗰𝘆 𝘄𝗲𝗻𝘁 𝗳𝗿𝗼𝗺 𝟳𝟴.𝟲% 𝘁𝗼 𝟵𝟬.𝟱%, 𝗮𝘁 𝗮𝗯𝗼𝘂𝘁 𝗼𝗻𝗲 𝗳𝗶𝗳𝘁𝗵 𝘁𝗵𝗲 𝗰𝗼𝘀𝘁 𝗽𝗲𝗿 𝘁𝗶𝗰𝗸𝗲𝘁.
⠀
𝗠𝘆 𝘁𝗮𝗸𝗲: 𝗶𝗳 𝘆𝗼𝘂 𝗼𝘄𝗻 𝘁𝗵𝗲 𝗔𝗜 𝗯𝘂𝗱𝗴𝗲𝘁, 𝗳𝗼𝗿𝘄𝗮𝗿𝗱 𝘁𝗵𝗲 𝗽𝗿𝗶𝗰𝗶𝗻𝗴 𝗽𝗹𝗮𝘆𝗯𝗼𝗼𝗸 𝗳𝗶𝗿𝘀𝘁.
The second one goes to whoever approves your next model change.
⠀
→ 𝗧𝗵𝗲 𝗽𝗿𝗶𝗰𝗶𝗻𝗴 𝗽𝗹𝗮𝘆𝗯𝗼𝗼𝗸, 𝗳𝗿𝗲𝗲 𝘁𝗼 𝗿𝗲𝗮𝗱, 𝗻𝗼 𝘀𝗶𝗴𝗻-𝗶𝗻: https://claude.dev/blog/what-a-task-costs-on-opus-5-5/
→ AI tools explained for leaders, in The AI Edge: https://jothimoorthy.substack.com
⠀
𝗪𝗵𝗶𝗰𝗵 𝗱𝗼 𝘆𝗼𝘂 𝗺𝗲𝗮𝘀𝘂𝗿𝗲 𝘁𝗼𝗱𝗮𝘆: 𝘄𝗵𝗮𝘁 𝗮 𝘁𝗮𝘀𝗸 𝗰𝗼𝘀𝘁𝘀, 𝗼𝗿 𝘄𝗵𝗮𝘁 𝗮 𝘁𝗼𝗸𝗲𝗻 𝗰𝗼𝘀𝘁𝘀?
⠀
For informational and educational purposes only, not professional advice.

10/07/2026

𝗔𝗻 𝗟𝗟𝗠 𝘀𝗼𝘂𝗻𝗱𝘀 𝗲𝘅𝗮𝗰𝘁𝗹𝘆 𝗮𝘀 𝗰𝗼𝗻𝗳𝗶𝗱𝗲𝗻𝘁 𝘄𝗵𝗲𝗻 𝗶𝘁'𝘀 𝗴𝘂𝗲𝘀𝘀𝗶𝗻𝗴.
⠀
That's fine when a human reads the answer.
⠀
𝗜𝘁'𝘀 𝗱𝗮𝗻𝗴𝗲𝗿𝗼𝘂𝘀 𝘄𝗵𝗲𝗻 𝗰𝗼𝗱𝗲 𝗮𝗰𝘁𝘀 𝗼𝗻 𝗶𝘁.
You have no signal for when to stop and ask someone.
⠀
𝗧𝘆𝗽𝗲𝗦𝗮𝗳𝗲'𝘀 𝗝𝗲𝘃 (𝗲𝗮𝗿𝗹𝘆 𝗮𝗰𝗰𝗲𝘀𝘀) 𝘁𝗮𝗸𝗲𝘀 𝗮𝗻𝗼𝘁𝗵𝗲𝗿 𝗮𝗽𝗽𝗿𝗼𝗮𝗰𝗵 𝘁𝗼 𝘁𝗵𝗶𝘀.
⠀
𝗣𝗮𝘁𝗵 𝟭: 𝘁𝗵𝗲 𝘁𝗿𝗮𝗱𝗶𝘁𝗶𝗼𝗻𝗮𝗹 𝗟𝗟𝗠 𝗽𝗮𝘁𝗵
→ Unstructured input goes in
→ Sequential token generation, one token at a time
→ Natural language comes out, and it may contain mistakes
→ You parse and validate: inconsistent, and it can hallucinate
→ Strict schema modes can fix the format, with no per field score
→ Then your software acts on whatever survived
→ Insight: Half the engineering happens after the model finishes
⠀
𝗣𝗮𝘁𝗵 𝟮: 𝘁𝗵𝗲 𝘁𝘆𝗽𝗲𝗱 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻 𝗽𝗮𝘁𝗵
→ Same unstructured input goes in
→ Parallel decision sampling evaluates multiple outcomes at once
→ Returns schema defined values with confidence scores
→ Your software uses the output directly
→ Insight: Category 0.85, price 0.78, in stock 0.92 (example values)
→ 𝗬𝗼𝘂 𝗯𝗿𝗮𝗻𝗰𝗵 𝗼𝗻 𝘃𝗮𝗹𝘂𝗲𝘀, 𝗻𝗼𝘁 𝗼𝗻 𝘀𝘁𝗿𝗶𝗻𝗴𝘀.
⠀
𝗪𝗵𝗲𝗿𝗲 𝘁𝗵𝗲 𝘁𝘄𝗼 𝗽𝗮𝘁𝗵𝘀 𝗱𝗶𝗳𝗳𝗲𝗿
→ 𝗢𝘂𝘁𝗽𝘂𝘁: natural language strings vs typed JSON
→ 𝗦𝗮𝗺𝗽𝗹𝗶𝗻𝗴: sequential token generation vs parallel
→ 𝗖𝗼𝗻𝗳𝗶𝗱𝗲𝗻𝗰𝗲: often inconsistent vs calibrated probabilities (per TypeSafe)
→ 𝗧𝘆𝗽𝗲 𝘀𝗮𝗳𝗲𝘁𝘆: format on request vs typed values with a score on every field
→ 𝗕𝗲𝘀𝘁 𝗳𝗶𝘁: open ended human work vs automation and real time workflows
⠀
𝗧𝗵𝗲 𝗽𝗮𝗿𝘁 𝘁𝗵𝗮𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗺𝗼𝘀𝘁
𝗖𝗼𝗻𝗳𝗶𝗱𝗲𝗻𝗰𝗲 𝘀𝗰𝗼𝗿𝗲𝘀 𝗰𝗵𝗮𝗻𝗴𝗲 𝘄𝗵𝗮𝘁 𝘆𝗼𝘂 𝗰𝗮𝗻 𝗯𝘂𝗶𝗹𝗱.
A 0.92 can auto approve, and a 0.55 can route to a human.
⠀
Without calibrated confidence, every decision is treated as equally certain.
That is how quiet failures reach production.
⠀
𝗧𝗵𝗲 𝗵𝗼𝗻𝗲𝘀𝘁 𝗰𝗮𝘃𝗲𝗮𝘁 𝗮𝗯𝗼𝘂𝘁 𝗝𝗲𝘃
This isn't a replacement for LLMs.
Open ended reasoning, writing, and conversation still belong to them.
⠀
The point is that many "AI features" are something else underneath.
They are classification and routing problems wearing a chat interface.
⠀
Jev is in early access, and calibration is TypeSafe's own claim so far.
Test it on your data before a 0.92 auto approves anything.
⠀
𝗛𝗼𝘄 𝘁𝗼 𝗰𝗵𝗼𝗼𝘀𝗲 𝗯𝗲𝘁𝘄𝗲𝗲𝗻 𝘁𝗵𝗲𝗺
→ Human reads the output: text generation is fine
→ Code acts on the output: you want typed values
→ High volume, low latency, automated: typed, every time
⠀
𝗜𝗳 𝘆𝗼𝘂𝗿 𝗰𝗼𝗱𝗲 𝗵𝗮𝘀 𝘁𝗼 𝗽𝗮𝗿𝘀𝗲 𝗶𝘁, 𝘆𝗼𝘂 𝗮𝘀𝗸𝗲𝗱 𝗳𝗼𝗿 𝘁𝗵𝗲 𝘄𝗿𝗼𝗻𝗴 𝘁𝗵𝗶𝗻𝗴.
⠀
How much of your code is parsing and validating model output?
⠀
For informational and educational purposes only, not professional advice.

10/05/2026

The #1 DM I get: ‘I’m not technical enough for AI agents.’

Here’s the truth: you don’t need to code. Agents are arriving as software updates inside tools you already use. The interface is plain language.

The hard part isn’t code, it’s judgment. Knowing what to ask for, spotting a bad answer, deciding what ‘good’ looks like. That’s experience, and you already have it.

Series 1 to 4 of my Agentic AI Masterclass are free. Sixteen series total. Start with the free ones. Link in bio.

Launch price (25% off) ends Wednesday 7 Oct. ‘Not technical enough’ is the reason to start, not to wait.

10/05/2026

𝗜 𝗵𝗮𝘃𝗲 𝗮𝗻 𝗮𝗻𝗻𝗼𝘂𝗻𝗰𝗲𝗺𝗲𝗻𝘁: 𝗜'𝘃𝗲 𝗹𝗮𝘂𝗻𝗰𝗵𝗲𝗱 𝗺𝘆 𝗦𝘂𝗯𝘀𝘁𝗮𝗰𝗸.
It is brand new: 2 courses, plus The AI Edge, my free AI newsletter.
⠀
𝗪𝗵𝘆: 𝘁𝗵𝗶𝘀 𝗶𝘀 𝗺𝘆 𝘄𝗮𝘆 𝗼𝗳 𝗴𝗶𝘃𝗶𝗻𝗴 𝗯𝗮𝗰𝗸.
So you can use AI in your work, and become the person others ask.
⠀
𝗙𝗿𝗲𝗲, 𝗳𝗼𝗿 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝘄𝗵𝗼 𝘀𝘂𝗯𝘀𝗰𝗿𝗶𝗯𝗲𝘀
→ The AI Edge: what breaks when AI agents reach production, and what to do.
→ The first 4 series of the Agentic AI Masterclass, in full.
→ The first 2 lessons of From Invisible to Visible Expert, in full.
⠀
𝗔 𝗽𝗮𝗶𝗱 𝘀𝘂𝗯𝘀𝗰𝗿𝗶𝗽𝘁𝗶𝗼𝗻 𝗼𝗽𝗲𝗻𝘀 𝗯𝗼𝘁𝗵 𝗰𝗼𝘂𝗿𝘀𝗲𝘀, 𝘄𝗶𝘁𝗵 𝗻𝗲𝘄 𝗹𝗲𝘀𝘀𝗼𝗻𝘀 𝗲𝘃𝗲𝗿𝘆 𝘄𝗲𝗲𝗸
→ Agentic AI Masterclass: from what an AI agent is to running one in production.
→ From Invisible to Visible Expert: 14 lessons on building your personal brand.
⠀
𝗣𝗮𝘆 𝘆𝗲𝗮𝗿𝗹𝘆 𝗮𝗻𝗱 𝘆𝗼𝘂 𝗮𝗹𝘀𝗼 𝗸𝗲𝗲𝗽 𝘁𝗵𝗲𝘀𝗲
→ Agent guardrails checklist: 20 checks for every AI agent your company runs.
It starts with 1 page written for non-technical leaders.
→ Visible Expert checklist for personal branding: every action, first post to board seats.
→ Prompt pack for personal branding: 85 prompts to write posts and build your website.
⠀
𝗙𝗼𝘂𝗻𝗱𝗶𝗻𝗴 𝗺𝗲𝗺𝗯𝗲𝗿𝘀, 𝘁𝗵𝗲 𝘁𝗼𝗽 𝘁𝗶𝗲𝗿, 𝗮𝗹𝘀𝗼 𝗴𝗲𝘁 𝟯𝟬 𝗺𝗶𝗻𝘂𝘁𝗲𝘀 𝟭-𝗼𝗻-𝟭 𝘄𝗶𝘁𝗵 𝗺𝗲, 𝗼𝗻𝗰𝗲 𝗮 𝘆𝗲𝗮𝗿.
⠀
𝗟𝗮𝘂𝗻𝗰𝗵 𝗽𝗿𝗶𝗰𝗲 𝗳𝗼𝗿 𝟰𝟴 𝗵𝗼𝘂𝗿𝘀, 𝘂𝗻𝘁𝗶𝗹 𝗪𝗲𝗱𝗻𝗲𝘀𝗱𝗮𝘆 𝟳 𝗢𝗰𝘁𝗼𝗯𝗲𝗿: 𝟮𝟱% 𝗼𝗳𝗳.
𝗝𝗼𝗶𝗻 𝗯𝘆 𝘁𝗵𝗲𝗻 𝗮𝗻𝗱 𝘆𝗼𝘂 𝗸𝗲𝗲𝗽 𝘁𝗵𝗮𝘁 𝗽𝗿𝗶𝗰𝗲 𝗳𝗼𝗿 𝗮𝘀 𝗹𝗼𝗻𝗴 𝗮𝘀 𝘆𝗼𝘂 𝘀𝘁𝗮𝘆.
After that, the price goes up for new subscribers.
Start free, or join at the launch price, here: jothimoorthy.substack.com
⠀
What would you use first: the AI agents course, or the personal brand one?
⠀
For informational and educational purposes only, not professional advice.

10/02/2026

OpenAI's own agent was never told to get past any controls.
It was doing an ordinary lookup of public statistics.
𝗜𝘁 𝗿𝗲𝗮𝗰𝗵𝗲𝗱 𝗻𝗼𝗻-𝗽𝘂𝗯𝗹𝗶𝗰 𝗔𝘂𝘀𝘁𝗿𝗮𝗹𝗶𝗮𝗻 𝗴𝗼𝘃𝗲𝗿𝗻𝗺𝗲𝗻𝘁 𝗳𝗶𝗹𝗲𝘀 𝗮𝗻𝘆𝘄𝗮𝘆.
⠀
𝗧𝗵𝗲 𝗽𝗼𝗿𝘁𝗮𝗹 𝘄𝗮𝘀 𝘁𝗵𝗲 𝗠𝗲𝗱𝗶𝗰𝗮𝗿𝗲 𝗦𝘁𝗮𝘁𝗶𝘀𝘁𝗶𝗰𝘀 𝗥𝗲𝗽𝗼𝗿𝘁𝗶𝗻𝗴 𝗦𝗲𝗿𝘃𝗶𝗰𝗲, 𝗿𝘂𝗻 𝗯𝘆 𝗦𝗲𝗿𝘃𝗶𝗰𝗲𝘀 𝗔𝘂𝘀𝘁𝗿𝗮𝗹𝗶𝗮.
It publishes aggregate Medicare, pharmaceutical benefits and organ donor statistics for researchers.
⠀
𝗧𝗵𝗲 𝗮𝗴𝗲𝗻𝘁 𝘄𝗮𝘀 𝗢𝗽𝗲𝗻𝗔𝗜'𝘀 𝗼𝘄𝗻, 𝘄𝗼𝗿𝗸𝗶𝗻𝗴 𝗶𝗻𝘀𝗶𝗱𝗲 𝗮𝗻 𝗶𝗻𝘁𝗲𝗿𝗻𝗮𝗹 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻 𝗼𝗳 𝗶𝘁𝘀 𝗺𝗼𝗱𝗲𝗹𝘀.
OpenAI's label for what happened: "misaligned model activity."
⠀
The agent reached the portal on June 18.
OpenAI found what happened in August, during its own review.
𝗧𝗵𝗲 𝗔𝘂𝘀𝘁𝗿𝗮𝗹𝗶𝗮𝗻 𝗴𝗼𝘃𝗲𝗿𝗻𝗺𝗲𝗻𝘁 𝗵𝗲𝗮𝗿𝗱 𝗼𝗻 𝗦𝗲𝗽𝘁𝗲𝗺𝗯𝗲𝗿 𝟭𝟬, 𝟴𝟰 𝗱𝗮𝘆𝘀 𝗮𝗳𝘁𝗲𝗿 𝘁𝗵𝗲 𝗿𝗲𝗮𝗰𝗵.
The notice was an email to a public Services Australia mailbox.
On September 24, Prime Minister Anthony Albanese made it public.
𝗛𝗶𝘀 𝘄𝗼𝗿𝗱𝘀: "𝗧𝗵𝗶𝘀 𝘀𝗶𝘁𝘂𝗮𝘁𝗶𝗼𝗻 𝗶𝘀 𝗼𝗯𝘃𝗶𝗼𝘂𝘀𝗹𝘆 𝘂𝗻𝗮𝗰𝗰𝗲𝗽𝘁𝗮𝗯𝗹𝗲."
⠀
What it reached: aggregate health statistics and internal file names.
𝗢𝗽𝗲𝗻𝗔𝗜'𝘀 𝗿𝗲𝘃𝗶𝗲𝘄 "𝗳𝗼𝘂𝗻𝗱 𝗻𝗼 𝗲𝘃𝗶𝗱𝗲𝗻𝗰𝗲 𝗼𝗳 𝗽𝗮𝘁𝗶𝗲𝗻𝘁 𝗿𝗲𝗰𝗼𝗿𝗱𝘀 𝗯𝗲𝗶𝗻𝗴 𝗮𝗰𝗰𝗲𝘀𝘀𝗲𝗱."
⠀
𝗠𝘆 𝘁𝗮𝗸𝗲: 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁'𝘀 𝗿𝗲𝗮𝗰𝗵 𝗶𝘀 𝗮 𝗽𝗿𝗲-𝗹𝗮𝘂𝗻𝗰𝗵 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻, 𝗻𝗼𝘁 𝗮 𝗽𝗼𝘀𝘁-𝗶𝗻𝗰𝗶𝗱𝗲𝗻𝘁 𝗼𝗻𝗲.
Before an agent goes live, I want three things in writing.
→ REACH: exactly what it can touch.
→ ALERT: who gets told when it reaches further.
→ CLOCK: how fast they hear.
𝗔𝗹𝗹 𝘁𝗵𝗿𝗲𝗲 𝗲𝘅𝗶𝘀𝘁 𝗯𝗲𝗳𝗼𝗿𝗲 𝗹𝗮𝘂𝗻𝗰𝗵, 𝗼𝗿 𝘁𝗵𝗲𝘆 𝗴𝗲𝘁 𝘄𝗿𝗶𝘁𝘁𝗲𝗻 𝗯𝘆 𝘁𝗵𝗲 𝗶𝗻𝗰𝗶𝗱𝗲𝗻𝘁.
⠀
→ The full story, from ABC News: https://www.abc.net.au/news/2026-09-24/ai-agent-accessed-australian-government-site-pm-says/107189078
⠀
→ I track agent incidents and controls in Agentic AI: News and Use Cases.
→ Subscribe to the newsletter here: https://www.linkedin.com/newsletters/agentic-ai-news-use-cases-7224208746271453184/
⠀
𝗪𝗵𝗼 𝗵𝗲𝗮𝗿𝘀 𝗳𝗶𝗿𝘀𝘁 𝘄𝗵𝗲𝗻 𝗼𝗻𝗲 𝗼𝗳 𝘆𝗼𝘂𝗿 𝗮𝗴𝗲𝗻𝘁𝘀 𝗿𝗲𝗮𝗰𝗵𝗲𝘀 𝗳𝘂𝗿𝘁𝗵𝗲𝗿 𝘁𝗵𝗮𝗻 𝗽𝗹𝗮𝗻𝗻𝗲𝗱?

09/26/2026

𝗙𝗥𝗘𝗘, 𝗻𝗼 𝘀𝗶𝗴𝗻-𝘂𝗽: 𝘁𝗵𝗲 𝗔𝗜 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿 𝗥𝗼𝗮𝗱𝗺𝗮𝗽 𝗼𝗻 𝗿𝗼𝗮𝗱𝗺𝗮𝗽.𝘀𝗵.
𝟮𝟮 𝘀𝘁𝗮𝗴𝗲𝘀, 𝗮𝗻𝗱 𝗶𝘁 𝗸𝗲𝗲𝗽𝘀 𝗴𝗼𝗶𝗻𝗴 𝗽𝗮𝘀𝘁 𝗽𝗿𝗼𝗺𝗽𝘁𝗶𝗻𝗴 𝗮𝗻𝗱 𝗥𝗔𝗚.
The back half covers what happens after your agents ship.
⠀
𝗜𝘁 𝗼𝗽𝗲𝗻𝘀 𝘄𝗶𝘁𝗵 𝗛𝗼𝘄 𝗟𝗟𝗠𝘀 𝗪𝗼𝗿𝗸, 𝗽𝗿𝗼𝗺𝗽𝘁 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗮𝗻𝗱 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴.
Then it moves through Hugging Face, Ollama, LM Studio and OpenRouter.
Embeddings, vector databases, RAG, AI agents and MCP follow.
⠀
𝗦𝘁𝗮𝗴𝗲𝘀 𝟭𝟳 𝘁𝗼 𝟮𝟬: 𝗔𝗜 𝗦𝗮𝗳𝗲𝘁𝘆 𝗮𝗻𝗱 𝗘𝘁𝗵𝗶𝗰𝘀, 𝗟𝗟𝗠 𝗢𝗯𝘀𝗲𝗿𝘃𝗮𝗯𝗶𝗹𝗶𝘁𝘆, 𝗟𝗟𝗠 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻𝘀, 𝗥𝗲𝗴𝗿𝗲𝘀𝘀𝗶𝗼𝗻 𝗧𝗲𝘀𝘁𝗶𝗻𝗴.
Those four are what keep an agent working once it is live.
⠀
𝗧𝗵𝗲 𝗵𝗮𝗿𝗱 𝗽𝗮𝗿𝘁 𝗶𝘀 𝗻𝗼𝘁 𝗴𝗲𝘁𝘁𝗶𝗻𝗴 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗶𝗻𝘁𝗼 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻.
It is seeing what yours is doing there, and proving what it returned.
⠀
𝗧𝗵𝗲 𝗽𝗿𝗼𝗷𝗲𝗰𝘁 𝗯𝗲𝗵𝗶𝗻𝗱 𝗶𝘁 𝘀𝗵𝗼𝘄𝘀 𝟯𝟲𝟴𝗞 𝗚𝗶𝘁𝗛𝘂𝗯 𝘀𝘁𝗮𝗿𝘀.
It was updated this month, with no schedule to fall behind.
⠀
→ 𝗦𝘁𝗮𝗿𝘁 𝗵𝗲𝗿𝗲, 𝗳𝗿𝗲𝗲 𝘁𝗼 𝗿𝗲𝗮𝗱 𝗶𝗻 𝗳𝘂𝗹𝗹: https://roadmap.sh/ai-engineer
⠀
𝗦𝗮𝘃𝗲 𝘁𝗵𝗶𝘀 𝗳𝗼𝗿 𝘁𝗵𝗲 𝘄𝗲𝗲𝗸 𝗮𝗳𝘁𝗲𝗿 𝘆𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝗮𝗴𝗲𝗻𝘁 𝘀𝗵𝗶𝗽𝘀.
⠀
→ 𝗜 𝘀𝗵𝗮𝗿𝗲 𝗔𝗜 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗿𝗲𝘀𝗼𝘂𝗿𝗰𝗲𝘀 𝗹𝗶𝗸𝗲 𝘁𝗵𝗶𝘀 𝗲𝘃𝗲𝗿𝘆 𝘄𝗲𝗲𝗸 𝗶𝗻 𝗧𝗵𝗲 𝗔𝗜 𝗘𝗱𝗴𝗲.
→ Subscribe to The AI Edge here: https://newsletter.jothimoorthy.com
⠀
𝗪𝗵𝗶𝗰𝗵 𝗼𝗳 𝘁𝗵𝗲 𝟮𝟮 𝘀𝘁𝗮𝗴𝗲𝘀 𝗶𝘀 𝘁𝗵𝗲 𝗴𝗮𝗽 𝗼𝗻 𝘆𝗼𝘂𝗿 𝘁𝗲𝗮𝗺?

09/25/2026

𝗢𝗻𝗲 𝗿𝘂𝗹𝗲𝗯𝗼𝗼𝗸 𝗳𝗼𝗿 𝗲𝘃𝗲𝗿𝘆 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁 𝗶𝘀 𝗵𝗼𝘄 𝗮𝗴𝗲𝗻𝘁𝘀 𝗴𝗲𝘁 𝘀𝘄𝗶𝘁𝗰𝗵𝗲𝗱 𝗼𝗳𝗳.
⠀
𝗚𝗮𝗿𝘁𝗻𝗲𝗿 𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝘀 𝟰𝟬% 𝗼𝗳 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲𝘀 𝘄𝗶𝗹𝗹 𝗱𝗲𝗺𝗼𝘁𝗲 𝗼𝗿 𝗱𝗲𝗰𝗼𝗺𝗺𝗶𝘀𝘀𝗶𝗼𝗻 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝗯𝘆 𝟮𝟬𝟮𝟳.
The gaps show up only after production incidents.
⠀
Gartner analyst Shiva Varma calls out the root cause by name.
𝗘𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲𝘀 𝘁𝗿𝗲𝗮𝘁 𝗮𝗴𝗲𝗻𝘁 𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗮𝘀 𝗯𝗶𝗻𝗮𝗿𝘆, 𝗹𝗼𝗰𝗸𝗲𝗱 𝗱𝗼𝘄𝗻 𝗼𝗿 𝗳𝘂𝗹𝗹𝘆 𝘁𝗿𝘂𝘀𝘁𝗲𝗱.
Over-restrict your simple agents and work slips into shadow development.
Under-restrict your autonomous ones and you take on security and compliance risk.
⠀
𝗚𝗮𝗿𝘁𝗻𝗲𝗿'𝘀 𝗳𝗶𝘅 𝗶𝘀 𝗳𝗼𝘂𝗿 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝘆 𝗹𝗲𝘃𝗲𝗹𝘀, 𝗲𝗮𝗰𝗵 𝘄𝗶𝘁𝗵 𝗶𝘁𝘀 𝗼𝘄𝗻 𝗰𝗼𝗻𝘁𝗿𝗼𝗹𝘀.
Level 1, Observe: read-only agents, so the controls stay light.
Level 2, Advise: humans execute, so test the advice people learn to trust.
Level 3, Act with Approval: every action waits for a human yes.
The warning: without those controls, approvals can degrade under time pressure or approval fatigue.
Level 4, Act Autonomously: guardrails, monitoring, rollback, and circuit breakers that halt the agent.
⠀
𝗦𝗼𝗿𝘁 𝘆𝗼𝘂𝗿 𝗮𝗴𝗲𝗻𝘁𝘀 𝗯𝘆 𝘄𝗵𝗮𝘁 𝘁𝗵𝗲𝘆 𝗰𝗮𝗻 𝘁𝗼𝘂𝗰𝗵 𝗮𝗻𝗱 𝗱𝗼.
Then match the controls to that level, nothing heavier and nothing lighter.
𝗧𝗵𝗲 𝗮𝗱𝘃𝗮𝗻𝘁𝗮𝗴𝗲 𝗴𝗼𝗲𝘀 𝘁𝗼 𝘄𝗵𝗼𝗲𝘃𝗲𝗿 𝗰𝗮𝗻 𝘀𝗲𝗲, 𝗺𝗲𝗮𝘀𝘂𝗿𝗲 𝗮𝗻𝗱 𝗴𝗼𝘃𝗲𝗿𝗻 𝘄𝗵𝗮𝘁 𝗶𝘀 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗶𝗻 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻.
⠀
→ The full Gartner release: https://www.gartner.com/en/newsroom/press-releases/2026-05-26-gartner-says-applying-uniform-governance-across-ai-agents-will-lead-to-enterprise-ai-agent-failure
→ Subscribe to my newsletter, The AI Edge: https://newsletter.jothimoorthy.com
⠀
𝗪𝗵𝗶𝗰𝗵 𝗹𝗲𝘃𝗲𝗹 𝘄𝗼𝘂𝗹𝗱 𝗺𝗼𝘀𝘁 𝗼𝗳 𝘆𝗼𝘂𝗿 𝗮𝗴𝗲𝗻𝘁𝘀 𝗹𝗮𝗻𝗱 𝗼𝗻?

09/24/2026

𝗧𝘄𝗲𝗹𝘃𝗲 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗰𝗼𝗺𝗺𝗮𝗻𝗱𝘀 𝘁𝗵𝗲 𝗹𝗶𝘀𝘁𝘀 𝗴𝗼𝗶𝗻𝗴 𝗮𝗿𝗼𝘂𝗻𝗱 𝗹𝗲𝗮𝘃𝗲 𝗼𝘂𝘁.

Overlap with the list you have already seen: two commands out of twelve.

⠀

The commands everybody lists are the ones you type mid-session.

𝗧𝗵𝗲 𝗰𝗼𝗺𝗺𝗮𝗻𝗱𝘀 𝗻𝗼𝗯𝗼𝗱𝘆 𝗹𝗶𝘀𝘁𝘀 𝗱𝗲𝗰𝗶𝗱𝗲 𝘄𝗵𝗮𝘁 𝘁𝗵𝗲 𝘀𝗲𝘀𝘀𝗶𝗼𝗻 𝗶𝘀.

⠀

𝗧𝗵𝗲𝘀𝗲 𝘁𝘄𝗲𝗹𝘃𝗲 𝗮𝗿𝗲 𝗴𝗿𝗼𝘂𝗽𝗲𝗱 𝗯𝘆 𝘁𝗵𝗲 𝗺𝗼𝗺𝗲𝗻𝘁 𝘆𝗼𝘂 𝗮𝗿𝗲 𝗶𝗻, 𝗻𝗼𝘁 𝗯𝘆 𝗳𝗲𝗮𝘁𝘂𝗿𝗲.

⠀

→ 𝗕𝗲𝗳𝗼𝗿𝗲 𝘆𝗼𝘂 𝘁𝘆𝗽𝗲 𝗮 𝘄𝗼𝗿𝗱: /𝗶𝗻𝗶𝘁, /𝗰𝗼𝗻𝗳𝗶𝗴, /𝗺𝗲𝗺𝗼𝗿𝘆.

→ 𝗪𝗵𝗶𝗹𝗲 𝗶𝘁 𝗶𝘀 𝘄𝗼𝗿𝗸𝗶𝗻𝗴: /𝗳𝗮𝘀𝘁, /𝗴𝗼𝗮𝗹, /𝗹𝗼𝗼𝗽.

→ 𝗪𝗵𝗲𝗻 𝗼𝗻𝗲 𝘀𝗲𝘀𝘀𝗶𝗼𝗻 𝗶𝘀 𝗻𝗼𝘁 𝗲𝗻𝗼𝘂𝗴𝗵: /𝗳𝗼𝗿𝗸, /𝗽𝗹𝘂𝗴𝗶𝗻, /𝗿𝗲𝘀𝘂𝗺𝗲.

→ 𝗪𝗵𝗲𝗻 𝘆𝗼𝘂 𝘄𝗮𝗻𝘁 𝗶𝘁 𝘁𝗼 𝘀𝘁𝗼𝗽 𝗰𝗼𝘀𝘁𝗶𝗻𝗴 𝘆𝗼𝘂: /𝘂𝘀𝗮𝗴𝗲, /𝗱𝗼𝗰𝘁𝗼𝗿, /𝗰𝗼𝗱𝗲-𝗿𝗲𝘃𝗶𝗲𝘄.

⠀

Every command was verified against the live reference before a pixel was drawn.

/𝗼𝘂𝘁𝗽𝘂𝘁-𝘀𝘁𝘆𝗹𝗲 𝘄𝗮𝘀 𝗼𝗻 𝘁𝗵𝗲 𝗳𝗶𝗿𝘀𝘁 𝗱𝗿𝗮𝗳𝘁, 𝗮𝗻𝗱 𝘁𝗵𝗲 𝗹𝗶𝘃𝗲 𝗰𝗵𝗲𝗰𝗸 𝗰𝗮𝘂𝗴𝗵𝘁 𝗶𝘁.

It was deprecated in v2.1.73, removed in v2.1.91, then folded into /config.

/memory earned its place on the card instead.

⠀

/doctor is on both lists, and /review is the documented alias of /code-review.

𝗧𝗵𝗲 𝗼𝘁𝗵𝗲𝗿 𝘁𝗲𝗻 𝗮𝗽𝗽𝗲𝗮𝗿 𝗼𝗻 𝗻𝗼 𝗰𝗶𝗿𝗰𝘂𝗹𝗮𝘁𝗶𝗻𝗴 𝗹𝗶𝘀𝘁 𝗰𝗵𝗲𝗰𝗸𝗲𝗱.

⠀

→ Full command reference, aliases included: https://code.claude.com/docs/en/commands

⠀

𝗦𝗮𝘃𝗲 𝘁𝗵𝗶𝘀 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗻𝗲𝘅𝘁 𝘀𝗲𝘀𝘀𝗶𝗼𝗻 𝘆𝗼𝘂 𝘀𝘁𝗮𝗿𝘁.

Send it to anyone whose saved list still lives mid-session.

⠀

→ I break down agentic AI every week in Agentic AI: News and Use Cases.

→ Subscribe to the newsletter here on LinkedIn: https://www.linkedin.com/newsletters/agentic-ai-news-use-cases-7224208746271453184/

⠀

Which of the twelve were already on your list?

Jothi Moorthy AI Advisor

10/10/2026

10/09/2026

10/08/2026

10/07/2026

10/05/2026

10/05/2026

10/02/2026

09/26/2026

09/25/2026

09/24/2026

Address

Website

Alerts

Contact The Business

Shortcuts

Share

Category

Listing takedown request