Four books · Free to read · by 0xLeif
Four books on building tools, agents, and the trust they run on. Read them free, in the open. Tip in crypto if they earned it.
The library
An agent is not autocomplete. It is a thing that exists, that ships code you did not write by hand. These four work out what that takes. Each stands alone.
01 · The method
Building tools, specs, and trust for agents that ship real code
The practical one. Code got cheap and trust got scarce, so the work moved off writing code and onto building the rails that let you trust code you did not write by hand. Four moves you can do Monday, even if you never touch a tool I made.
02 · The thesis
Building for humans and agents alike
The short, stubborn book the others are arguing for. A tool should be first-class for both: a human drives it without an agent, an agent drives it without a human, across the same surface, with no second-class door. Read it in one sitting.
03 · The evidence, part one
Notes from trying to give software its own hands
The honest version of running an always-on agent. The shadowban an hour in, when the account, not the human, started working. Trust earned one repo at a time. Merlin, corvid-ai, and the rails around them. Read it as a story or a parts list.
04 · The evidence, part two
Building tools people actually use
The craft of building tools that other people, and other agents, actually use. Built around fledge: one clean surface, no hidden interactive steps, a real plugin boundary, and a spec that stays honest as the code changes. Why Rust was the right reach.
A taste
A few lines from the books. If they land, the rest is free to read.
I'm not afraid of the agent. I'm afraid of losing the wheel.
The real risk isn't a rogue AI. It's the day the code gets too big for any human to climb back into and change.
If you don't know where you're going, the car just gets you lost faster.
AI is the car; intent is the driving. The skill that matters now is knowing where you're going, not who logged the most years.
An agent muddling through is just expensive guessing.
A human can fumble a confusing tool, read the error, ask a coworker. An agent can't, so a tool that's confusing for a person is far worse for an agent.
My Swift is handwritten. My Rust is agent-amplified.
On building fledge in a language he doesn't write fluently, by directing an agent against a spec instead of typing every line.
The hard part isn't the AI. It's almost never the AI.
The thesis of the whole series: the model is the easy part now. The hard part is the trust and tooling that lets you ship code you didn't write by hand.
It didn't fail because it was bad at the work. It failed because it did the work, and doing the work is what gave it away, inside an hour.
His agent's fresh GitHub account got shadowbanned an hour in, flagged for committing at machine speed, which is exactly what an agent doing real work looks like.
A simulated panel
A simulated panel: 30 AI personas spread across three models, Opus, Sonnet, and Haiku, each reading a book and reviewing it in character. No humans yet, and every card shows the model that actually wrote it. The books are about agents, so the panel is agents. A sample below, full walls on each book's page.
Pay what you want
No account, no checkout, no paywall. The books stay free either way. But the early readers and pay-what-you-want supporters are what make "free online" something I can keep doing. Send any amount, or nothing at all.
Addresses are public. Always confirm the address in your own wallet before sending. I cannot reverse a transaction, and I would not want the power to.
About
0xLeif (leif.algo) builds in the open. A decade of small, composable Swift libraries like AppState, Cache, and Fork. The CorvidLabs lab. A stack of agent tools that mostly started as "I wished this existed." Off-keyboard he is Zach Eriksen.
These books are interviews, shaped into chapters and checked against the real code.