Software factories: an X thread, against our ledger

What we read

The replies in the thread, and the pieces they leaned on: StrongDM’s software factory, OpenAI’s “Harness engineering”, Anthropic’s “Harness design for long-running application development” and “Demystifying evals for AI agents”, Dan Shapiro’s “The Five Levels”, Steve Yegge’s “Welcome to Gas Town”, and Addy Osmani’s “The Factory Model” and “Software Factories, Light and Dark”.

What our ledger said

Over five days, 8 to 13 October, building took 6.6 of 121.7 hours, about 5% of the time, and 54 of 59 cards built on the first attempt. So neither the number of agents nor their freedom was the limit. What the ledger cannot say is what the other 95% waited on: deciding and specifying the work, deliberate pacing, or the owner’s own hours.

The reading we work from

Jaka Hudoklin’s: writing the prompts and saying what should be different takes most of the time, and imagination is the bottleneck, not the agents. Our numbers fit his reading, but they fit pacing and the owner’s hours too, so it stays a working reading, not a finding.

So the kit keeps one build slot until a queue forms, and the effort goes into the spec: every medium or large card Claude writes is read by a second model family before it is built, unless you waive the read.

Where that puts us

Most replies sat on one line, from one person steering one agent to factories that run with nobody watching. In Igor’s phrase from the thread, we sit at harnessed parallelism, with one slot and a criticality dial: many cards, one at a time, a person at the plan, and installs that wait for your key unless you pre-approve them.

The tags it produced

• kept Intent is the constraint

All the tags.

Sources

• StrongDM’s software factory
• OpenAI, “Harness engineering”
• Anthropic’s engineering posts
• Dan Shapiro, “The Five Levels”
• Steve Yegge, “Welcome to Gas Town”
• Addy Osmani’s essays

A credit names where an idea came from. None of these people has endorsed the kit.