2AM Logic is one engineer and a fleet of AI agents. That sentence usually reads as either a boast or a confession, so this first note is about what it actually means, mechanically.
The loop
Work gets queued as issues, each carrying a spec and the criteria a reviewer should hold it to. Agents claim the issues, implement, review one another's pull requests, and merge, coordinated through the same forge a human team would use. The orchestration layer is Loom, a harness we built that turns issue labels into a claim-implement-review protocol, so the coordination lives in the repo itself rather than in anyone's head. The merge queue does not care what time it is.
What comes back is a pile of merged work and a pile of questions. The human job is judgment: deciding which of those merges was actually right, and tightening the specs where the answer was no.
The evidence
vibesql is a SQL:1999 database in Rust with every line written by agents. It passes 100% of SQLLogicTest, about 7.4 million tests, and separately reached full Core conformance in under two weeks of wall-clock time. That last number is the one that matters here: one engineer reviewing, nobody else writing Rust, under two weeks. Whatever a database engine used to cost in labor, it does not cost that anymore.
kicad-tools points the same loop at hardware. Agents take a design from schematic through a routed PCB to a downloadable fabrication package, and the gallery shows the boards to prove it, from a simple LED board up through an STM32 devboard and a brushless motor controller.
What the human does
Specification is the real work now; implementation is the easy part. The engineering that matters is deciding what to build, encoding how to verify it, and building the rigs (the orchestration, the review gates, the test oracles) that let agents carry a spec to a merge without a human watching each step. Judgment stays human. The typing, increasingly, does not.
Why this points at hardware
If software development costs are collapsing, the interesting question is what becomes cheap next. Our bet is hardware engineering. The tools are older, the feedback loops are slower, and almost nobody has made them agent-native, which is exactly what makes the gap worth working in. We are starting with electronics and moving down the stack toward chip design, in the open.
What breaks first when you point a fleet of agents at a chip design flow? We intend to find out and write it down here. (And yes, the agents do a lot of it at 2AM. They don't know what time it is.)
Update, 2026-08-04
This note never sized its fleet. Over the seven days from 2026-07-28, the day it went up, the loop merged 488 pull requests across the eight public repositories in our GitHub org, about seventy a day. The window is closed, so the count does not move: org:2AMLogic is:pr is:merged is:public merged:2026-07-28..2026-08-03. The is:public qualifier is load-bearing, because without it an authenticated org member sees 506 and picks up repositories this note never claimed.
Moving down the stack toward chip design has artifacts as of 2026-08-04: klayout-tools, which drives layout, DRC and LVS, plus seven public block repositories designed against it on the open PDKs, from a bandgap reference to a 10-bit SAR ADC.
The closing question above has an answer. Two bandgaps, in the open, three days after this note, reported what broke first: the spec did. The agents had specified a 1.20 V reference to ±1% untrimmed, and a review pass ran the repo's own device mismatch measurements through the amplifier gain, showed ±1% was contradicted by them, and moved the number to ±2%. That is the part software never had to do. vibesql never had to write its own exam, because SQLLogicTest already existed, 7.4 million cases of it, free from the first commit. Nothing in hardware hands you the spec suite, so somebody has to build the thing that says pass or fail, which is the harder problem and also the only reason you find out.