Posts

If 2025 is the year of agents, then 2026 will surely belong to agent managers.

Agent managers are people who can manage teams of AI agents. How many can one person successfully manage?

I can barely manage 4 AI agents at once. They ask for clarification, request permission, issue web searches—all requiring my attention. Sometimes a task takes 30 seconds. Other times, 30 minutes. I lose track of which agent is doing what & half the work gets thrown away because they misinterpret instructions.

For the last decade, the biggest line item in any startup’s R&D budget was predictable talent. But AI is pushing its way onto the P&L.

How much should a startup spend on AI as a percentage of its research and development spend?

10%? 30%? 60?

There are three factors to consider. First, the average salary for a software engineer in Silicon Valley. Second is the total cost of AI used by that engineer. Cursor is now at $200 per month for their Ultra Plan & reviews of Devin suggest $500 per month. Third, the number of agents an engineer can manage.

When you query AI, it gathers relevant information to answer you.

But, how much information does the model need?

Conversations with practitioners revealed the their intuition : the input was ~20x larger than the output.

But my experiments with Gemini tool command line interface, which outputs detailed token statistics, revealed its much higher.

300x on average & up to 4000x.

If I have a dollar to invest in a stock or a crypto token, how do I decide? I need to compare across the two.

Historically, that comparison was impossible. Crypto traded on a potent cocktail of hype, narrative, & the promise of a decentralized future. Perception drove valuations.

That’s changing. The word “revenue” is no longer verboten in the world of crypto. It’s becoming the goal.

This trend will unlock the next wave of institutional capital because investors can compare the risk/reward of crypto with the same metrics as other software companies.

Yesterday, Figma filed its beautifully designed S-1.

It reveals a product-led growth (PLG) business with a remarkable trajectory. Figma’s collaborative design tool platform disrupted the design market long-dominated by Adobe.

Here’s how the two companies stack up on key metrics for their most recent fiscal year:

Figma is about 3% the size of Adobe but growing 4x faster. The gross margins are identical. Figma’s 132% Net Dollar Retention is top decile.

Gmail’s AI email assistant writes like a committee of lawyers designed it.

Pete Koomen’s recent post Horseless Carriages explains why: developers control the AI prompts instead of users. In his post he argues that software developers should expose the prompts and the user should be able to control it.

He inspired me to build my own. I want a system that’s fast, accounts for historical context, & runs locally (because I don’t want my emails to be sent to other servers), & accepts guidance from a locally running voice model.¹

In “Data Rules Everything Around Me,” Matt Slotnick wrote about the difference between SaaS & AI apps. A typical SaaS app has a workflow layer, a middleware/connectivity layer, & a data layer/database. So does an AI app.

AI makes writing frontends trivial, so in the three-layer cake of workflow software the data matters much more.

The big differences between an AI & the SaaS app lie within the ganache of the middle layer. In SaaS applications, coded business rules determine each step a lead follows from creation to close.

Remember when you took a family photo & Ghibli-styled it?

Or that vibe coding session, when you pasted a screenshot of the browser so the AI can help you debug some Javascript?

Today, we expect AI to be able to hear, see, & read. This is why multimodal is the future of AI.

Multimodal data means using text, images, video, sound, even three-dimensional shapes with AI.

Systems of record are recognizing they cannot “take their survival for granted.”

One strategy is to acquire : the rationale Salesforce gives for the Informatica acquisition.

Another strategy is more defensive - hampering access to the data within the systems of record (SOR).

Unlike the previous software era where SORs built platforms on top of themselves to develop broader ecosystems (in Salesforce’s case Veeva & Vlocity), the AI shift does seem to be more defensive.

DuckLake is one of the most exciting technologies in data.

While data lakes are powerful, the formats that manage them have become notoriously difficult to work with.

“I think one of the things in DuckLake that we managed to do is to cut, I want to say like 15 technologies out of this stack.”

How does it achieve this? Instead of building a custom catalog server, DuckLake uses a simple, elegant idea: a standard database to manage metadata. It uses a database for what it’s good at. This clean architecture allows DuckLake to manage huge data lakes—with millions or even billions of files—across AWS S3 or Google Cloud Storage.