Notes from the operator seat.
Essays on the operating model behind Gigaverse — what worked, what I got wrong, and the systems underneath. Free to read, no email wall.
Protocols
Anthropic gave MCP away to a foundation whose platinum members include every one of its competitors. Then the spec was rewritten to run behind an ordinary load balancer. You can watch a protocol become infrastructure by reading its changelog.
The services-as-software land grab
AWS put $1B into forward-deployed engineering and Microsoft $2.5B in five weeks. When the hyperscalers move into your thesis, did the door close — or did they just prove the room is worth entering?
The new lifecycle
Every phase of the software lifecycle exists because writing code was expensive. That assumption is what changed — and an industry that spent twenty years abolishing the upfront specification is quietly rebuilding one.
Code review
Over seven months, reviewers of AI-written pull requests approved more and commented less. The researchers' reading is not that the code got better — it is that the reviewers got tired.
Machine readable
This site had a file whose entire job was to tell AI systems what I had written. It listed three essays. I had written fifteen — and nothing broke, which is the actual lesson. Plus the checklist: what to fix, what the research says works, and how to verify it yourself.
Taste
Ira Glass described a gap between your taste and your ability, and said you close it by working. Agents closed it from the other side, for everyone, at once — and the scarce input is now the one that was never scarce before.
Selling outcomes, not hours
When agents do the work, both the seat and the billable hour stop making sense. How an AI-native firm actually prices — and why every step toward outcome pricing moves risk onto your own balance sheet.
The fleet
An open-source agent went from nothing to 247,000 GitHub stars and five names in about fourteen weeks. NVIDIA's response was not a better agent — it was a sandbox, which tells you what kind of problem this turned out to be.
Graph engineering
On the benchmarks everyone cites, knowledge graphs are losing to plain vector search — cheaper, simpler, often more accurate. Why I still think the memory substrate for a production agent is a graph.
Prompt injection
The industry keeps treating a security problem as a prompting problem. A mitigation that lives inside the token stream is graded by the attacker, not by you — and the only defences with a track record assume the model will be fooled.
Agentic engineering
Sixteen experienced developers were 19% slower with AI tools — and believed afterwards they had been 20% faster. Everything I know about this craft starts with that gap.
Harness engineering
Swapping the model changes less than you think. In a controlled comparison, the software wrapped around the model moved results 7.8× more than the model did — and that software is the part you actually own.
The unglamorous wedge
Everyone is building a horizontal copilot. The durable money is going into boring verticals nobody names at a dinner party — like a dental front desk that quietly misses one in three new-patient calls.
Your next customer is an agent
In one holiday season, AI-referred retail traffic jumped nearly 700% and the checkout button moved off the store into the chat. When the buyer is an agent, your catalogue — not your storefront — is the product.
Building from Dhaka
Almost everything written about building AI-native companies assumes a San Francisco zip code. Why building a studio from Bangladesh — with a hub in Wyoming — is an advantage in the agent era, not an apology.
The operator's stack
The frontier everyone points at is the one-person billion-dollar company. I'm building the thing next to it: one senior team running a portfolio of companies on shared machinery and a lot of agents.
What survives production
95% of enterprise AI pilots show no measurable P&L, and Gartner expects 40%+ of agent projects scrapped by 2027. Almost none of it is a model problem — it's a loop problem, and the loop is buildable.
Loop engineering
Agent capability has doubled roughly every seven months for six years. Reliability has not kept pace. The discipline that closes the gap is the loop you build around the model.
Software factories
Three engineers, a million lines, no hand-written code. What the agentic software factory actually is, which claims the evidence supports, and where it quietly breaks.
Services become software
The $200B SaaS pool was never the prize. What a year of building into the $4.6T services market taught the market — and why implementation became the moat.
The AI Operator
Every essay goes out to the list first. A dispatch on building AI-native companies, from inside Gigaverse.