Chamath - Software Factory

Coding agents, agent orchestration dashboards, and AI benchmarks each fail at least one of the five factory tests and are more accurately categorized as developer tools, not software factories.

When production costs collapse, value migrates to whoever can guarantee output — as has happened in every prior industrialization wave — meaning accountability, not code generation speed, will define the winning software factory.

When evaluating any vendor calling itself a software factory, ask one question: 'When the system breaks in production, who takes the call?' — the answer must be 'we do.'

Ask vendors to demonstrate live, on a real system, that changing a requirement propagates to code and that a hotfix propagates back to the requirement — the full coherence loop closed end-to-end.

8090's Software Factory reverse-engineered an 18-million-line COBOL and Assembly billing engine into more than 100,000 plain-English rules in 40 days.

Hitachi opened 'Software Works' in 1969 as a literal factory using statistical quality control, measuring defect rates per thousand lines of code — the first known software factory.

In 2004, two Microsoft architects published 'Software Factories,' arguing software should be built like cars: from proven components on repeatable production lines.

Code generation accelerates architectural drift: if agents produce 10x more code against unsynchronized specifications, the 18-million-line legacy problem that took four decades to build by hand could be replicated in a few years.

A publicly traded health insurer using 8090's deterministic pre-filter cut claims routed to a pay-per-catch vendor by more than 80%, avoiding more than $20 million over four years.

Most AI tooling contracts disclaim accountability for output in a single sentence, stating that output is provided as-is and verification is the customer's problem.

Writing new code was never the enterprise bottleneck; the bottleneck is that real systems are changed weekly by dozens of people, and every change risks requirement-to-documentation-to-code drift that compounds over decades.

Regulated industries have been slowest to adopt AI coding tools because 'the model wrote it' is not an answer an auditor accepts — audit trails must be intrinsic to production, not retrofitted.

Systems that require a 'hero' individual to function have neither true accountability nor factory-level reliability; heroes eventually leave, taking institutional knowledge with them.

Test 1 — Business Intent as Input: A factory's input must be business requirements expressed in business language, not engineering tickets written engineer-to-engineer.

Test 2 — Coherence Under Continuous Change: A factory must keep intent, specification, code, tests, and production behavior synchronized as a single governed object; changing a requirement must propagate to code and vice versa.

Test 3 — Person-Independence: A factory must produce at predictable speed and quality regardless of who is on shift; knowledge must compound in the system, not in individuals.

Test 4 — Full Traceability: Every unit of output must carry provenance — this rule exists because of this requirement, approved by this person, implemented in this change, verified by this test, deployed at this time — generated as a byproduct of production, not written after the fact.

Test 5 — Accountability for Finished Product: The factory vendor must accept responsibility when the system breaks in production — fixing it and absorbing the cost — not transfer all verification to the customer.

“Ford never sold you a wrench, some parts and wished you good luck. Ford sold you a car, and if the car failed, Ford recalled it, because it was their factory that produced it.”

“A system that needs a hero to function has neither accountability nor a factory. It has a hero, and heroes eventually find new adventures.”