Tony Yang

Founder, MindRocket AI: done-for-you AI agent teams for SME. 1:1 AI coach. 20 years in business.

[Tony Yang @TonyYangAI]

Right AI for the right job:

Image or flyer: ChatGPT
Video you don't have time to watch: Gemini
Long document, key points fast: Claude
Facts with sources: Perplexity
Excel, Word, Outlook: Copilot
Talk it through in the car: Grok voice
Whole job handed off: an agent

[Tony Yang @TonyYangAI]

BREAKING: Anthropic, maker of Claude, filed to go public. It wants a Nasdaq listing as "ANT."

What its IPO filing tells owners using AI agents:

1. Anthropic committed to spend at least $110.0B with Amazon through April 2036, "regardless of actual usage"

[Tony Yang @TonyYangAI]

My take: agentic AI is now mainstream. The agentic AI race just got more real, and faster.

If you want to catch up, let's talk.

https://t.co/92iK8vbTzB

[Tony Yang @TonyYangAI]

6. Over 90% of the contributions to Anthropic's own codebase in September 2026 were written by Claude

7. It launched Ode with Anthropic to bring Claude "to mid-sized companies across sectors"

[Tony Yang @TonyYangAI]

@5amuelchan One file every agent reads first is a clean way to change the mix in one place. That many changes that fast says the file is earning its keep. Do you note the reason next to each switch, or only the current pick?

[Tony Yang @TonyYangAI]

@chrisralarcon Approving the storyboard first puts the gate where it's cheapest. Sixteen loops means the agent never had a clear picture of done. One human yes up front beats a fix loop with no exit.

[Tony Yang @TonyYangAI]

@0xNotMarc Good find. The part I'd underline for banks is where the hard limits go: approval gates and per-user permission checks move into the tools, so one agent can serve people with different access.

[Tony Yang @TonyYangAI]

@seanmm333 Opening every link is the step almost nobody does. The claim with no link is the one I'd watch. Did it turn out true when you checked it yourself, or is that where these answers usually drift?

[Tony Yang @TonyYangAI]

@LouieMota88 Read-only for the reviewer makes sense. A checker that can also edit tends to fix things quietly instead of flagging them. For the agents that do write, I'd keep anything that leaves the building behind a human yes.

[Tony Yang @TonyYangAI]

@tryapexweb "Nobody reads log files" says it all. Making a test reply that reaches you part of the definition of done is the right fix. A yes from a buyer is exactly the event that should never land somewhere a person won't look.

[Tony Yang @TonyYangAI]

@DrAliDaniel Good habit, and an easy one to teach family. Dates of service and the provider's name can point back to someone too, so I'd blank those when the charge amounts are what you need explained.

[Tony Yang @TonyYangAI]

@LewisValeLV Memory over speed is the right ranking. Asking a client the same thing twice costs more trust than a slow reply. Reading everything before it goes out is what makes that memory safe to rely on. Who gets to correct a note when it's wrong?

[Tony Yang @TonyYangAI]

@BenjieMalinao Good pick. Making things is easy to hand off. Keeping track of what you told people you'd do is where trust is won or lost, and it's the job nobody notices until it slips.

[Tony Yang @TonyYangAI]

@chijuano The BPO playbook carries over well. Writing it down for the AI also shows you which steps only ever lived in one person's head. That gap is worth finding even before an agent touches the work.

[Tony Yang @TonyYangAI]

Sources: OpenAI, Projects in ChatGPT: https://t.co/mHfDm1jxou ; Anthropic, Claude Projects: https://t.co/imrmKGZlvD

[Tony Yang @TonyYangAI]

Ouch. One ChatGPT chat for your whole business mixes every job together. Give each job its own project.

When your pricing questions and a customer complaint share one thread, every new question carries all that old context with it.

Fix it in ChatGPT Projects or Claude Projects:
1. Create one project per job, such as "Sales replies" and "Content."
2. Add instructions to each one. In ChatGPT: the ... menu, then Project settings. In Claude: Set project instructions. Example for Sales replies: "Answer like our front desk. Quote only from the attached price list."
3. Upload only the files that job needs. Your price list goes in Sales replies.
4. Add one short page of facts that never change, like what you sell and who you sell to, to every project.
5. Start a new chat inside the project for each new task. The project keeps the files and instructions, so the chat stays on one job.

Today: open ChatGPT, click New project, and name it after the job that fills most of your old chat.

[Tony Yang @TonyYangAI]

Here's Grok Bot running a real workday on your phone, in 18 seconds.

[Tony Yang @TonyYangAI]

Holy cow.

Grok Bot can now use the apps on your phone. The update rolled out Oct 22 on iOS and Android.

It can reach six apps, and each one stays off until you tap Allow:
1. Calendar: see your day and add events
2. Reminders: set and check reminders
3. Contacts: find and add contacts
4. Messages: read and send messages
5. Location: find places near you
6. Home: control lights, locks and more

How a business owner can start this week:
1. Allow Calendar and Reminders first. Say "add a call with the supplier Tuesday at 10" while you drive.
2. Leave Messages off until you trust it. On iPhone, iMessage is not part of this update.
3. Ask for tomorrow's schedule at the end of each day and fix what it gets wrong.
4. Test it on small tasks for a week before you hand it anything a customer will see.

Early testers report mixed results, so check its work.

Save this for when the update reaches your phone.

[Tony Yang @TonyYangAI]

@TheAceWheelie Good takeaway for owners. When a vendor says a tool was tested, the useful follow-up is what was left out of the test. A clear answer to that usually tells you more than the result itself.

[Tony Yang @TonyYangAI]

@farrux_hewson Naming which account it used on every action is the part I'd care about most. I run a work and a personal Gmail through my own business assistant too, and a wrong sender is the kind of mistake you want caught before it goes out. Does it ask when it can't tell?