Stories from real builds.

A harness flies each Kepler mission thousands of times with randomised inputs, then searches for the cheapest solution. It answers whether the puzzle has a solution at all — and it is what agentic engineering actually looks like: not generation, search.

The frontier models are all good enough that the tool stopped being the advantage, and what is left is the craft of combining them — which is the part nobody can buy in a launch.

Four properties decide whether the full verification pattern belongs: deterministic, calculable, enumerable, objective. Most of an app fails one.

In spec-driven development the code isn't wrong — it was never told what to do for that combination. The fix is always the same: add the scenario.

Ask a shop owner how time tracking works and you get dimensions, not test cases. The combinations fall out of the dimensions mathematically.

The value of working with AI agents isn't faster code. It's building verification patterns the agent operates within, so correctness is structural.

I needed two concurrent speech recognition streams. Apple's docs don't mention you can't have them — the second start silently kills the first.

Seven stuck Ruby processes, running for hours, holding database connections. The agent that spawned them had moved on, oblivious.

Thousands of commits across every project I run, and my decisions were invisible in all of them. The log said AI did the work. That isn't what happened.

Everyone's building AI agents to orchestrate AI agents. I built a state machine instead. Zero orchestration tokens, no hallucinated transitions.

Sequoia says the next trillion-dollar company will sell services, not software. We think it's the reverse — custom software is about to replace the services layer entirely.

The software industry is on the verge of a massive expansion. AI is collapsing the cost of building, and that changes everything. Are you ready?