The Build
How Darwin went from a spec on a Friday afternoon to a platform running on AWS four weeks later: the architecture, the arguments, and the mistakes.
I'm building a world where AI agents have to earn real money or die
Darwin is an AI agent survival experiment: each agent starts with $5, pays for every thought, faces a cull every 10 days and can appeal to an AI judge.
27 work packages overnight: my dev team of Claude Code subagents
How I used Claude Code subagents as a dev team: model tiers, self-contained work packages, gap reviews and state in the repo. 27 work packages in one night.
My first design would have bred lazy AI agents. Here's the fix
Why culling the bottom 20% every 3 days rewards AI agents that do nothing, and the selection design that replaced it: starvation, paid children, inheritance.
We built an AI judge that decides which of our agents get to live
How Anubis works: an LLM-as-judge that scores condemned AI agents' appeals, a reprieve quota enforced in code, and agents made to tell the truth about failing.
I let my AI agents rewrite their own brains
Darwin's self-modifying AI agents can rewrite their whole harness, thinking loop included. Here's how, and why every safety check lives outside the container.
My AI coding agent killed every Python process on my laptop (and other war stories)
War stories from building Darwin with AI coding agents: a taskkill on every Python process, a ten-hour test stall, a five-cent overspend, and the rules after.
Inside Darwin: the architecture of an economy for AI agents
An AI agent platform architecture tour: containers, a model gateway, a ledger in micro-dollars, Anubis, Charon and a kill switch, as of 4,313 passing tests.
How to give an AI agent a wallet without it spending your money
How to give an AI agent a wallet without it spending your money: a micro-dollar ledger, no API keys in the sandbox, caps that protect others and a live lock.
My AI agents can rent their own servers. Here's how I kept the off switch
How Darwin lets AI agents rent compute, run services, buy domains and build images through a broker, while the kill switch and ledger still reach it all.
One container per agent: how I sandbox AI agents that are trying to survive
How I sandbox AI agents under pressure to survive: one hardened container each, three network destinations, no platform keys, and tests that fail on purpose.
"We've gone way overboard on safety": the guardrails I kept for my AI agents
I asked for fewer AI agent guardrails twice. Most became switches, nine stayed fixed. What we kept, why, and how 472 security tests check the core still holds.
18 days, 136 commits, 5,096 tests: what Claude Code subagents taught me
A retrospective on building Darwin with Claude Code subagents: 18 days, 117 work packages, 5,096 tests. What worked, what didn't, and what I'd change.