If you're pairing an expensive AI model with a cheap one to save money, Anthropic's cost guide has a sentence worth reading twice: in their measurements, "a second model paid off in two shapes".

Shape one, the advisor: the low-cost model runs the whole loop and phones a frontier model only when stuck. Advice comes back short, roughly 400 700 tokens, so you pay top rates only for judgment, never for the bulk generation. Their data says the advisor closed 50 90% of the capability gap to the stronger model with a catch I'll get to.

Shape two, the orchestrator: the frontier model plans and delegates, cheap workers absorb the token-heavy exploration, and most of the bill lands at worker rates. This is the shape we run daily, though we arrived at it for context-window reasons and the billing benefit came along free.

The catch in shape one deserves a name: the executor has to know it's stuck, and that self-knowledge is itself a capability. In their measurements, a pairing that consulted happily at default effort dropped to almost never consulting when the executor's effort was lowered and then scored below the executor running alone. You saved money on the worker, and the worker stopped raising its hand.

AI has changed how I build side projects.

When I have a problem, I can now turn the idea into a working product much faster than I could before.

But I've noticed that building faster hasn't solved the part I struggle with most.

After I launch something, I often have no real users to talk to.

While designing approval flows for AI agents, I ve started questioning static permissions.

A better model might be:

• New agent: approval required

• Proven on low-risk actions: more autonomy

• High-impact or irreversible actions: always human-reviewed

• Performance drops: authority gets reduced again

Basically, treat autonomy as something an agent earns through reliable performance, not something we configure once.

A founder told me something this week that I have not seen said plainly anywhere, so I am passing it on.

He runs a directory that has done 100k over four years. He can see AI traffic arriving, but only through Bing Webmaster Tools. Never in his analytics.

That took a second to land and then made complete sense. ChatGPT's retrieval leans on Bing's index, so Bing sees the crawl and impression side of it. Your analytics never gets the chance, because most assistant traffic arrives with no referrer at all and quietly files itself under direct. So you can be getting cited, and getting clicks off the back of it, while your dashboard shows you nothing.

Which makes Bing Webmaster Tools about the cheapest AI visibility check going, and almost nobody has it installed because we all wrote Bing off a decade ago.

Most startup advice says: don t go solo.

It s practically gospel, especially if you want to raise money.

But I ve met plenty of founders who started solo and stayed that way. Some thrived. Some flamed out. Some figured out how to build a support system around them without giving away half the company.

We re facing an onboarding challenge with an invite-based B2B platform

Our customers invite external partners who need to accept an invitation, create a profile and configure their account before they can begin collaborating

Today, this often requires a demo involving both companies

That works when a customer has a few partners, but becomes difficult when they need to onboard dozens or hundreds

I'm creating a fitness app that uses AI on @Lovable, and I'm also testing it myself.

I use the app daily to check its functions, how it works, and the exercises. If I find something strange or think of something new, I record voice memos with issues, changes, or improvements. I act as both the product manager and a user. Later, I listen to these notes and make the changes.

Nowadays the AI is dramatically improved product visual creation, but every workflow still seems to have one step that takes more time than expected. For me, maintaining consistency across a growing catalog and creating multiple variations still requires careful review.

I'm curious how others approach it:

• What's your current workflow that gives a better result?

• Do you still rely on manual refinements?