Factory on Building a Software Factory: 16-Hour Missions, 50% Fewer Tokens, Sequential Agents
AI Engineer · youtube · 2026-09-27
On the AI Engineer podcast, Factory's Tereza Tížková explains what it takes to run software factories for enterprises like EY and Adobe — the full software lifecycle run autonomously: collecting signals, prioritizing, building, validating and improving. Writing code, she argues, is the easy part; a software factory is not just a swarm of coding agents.
Three principles and key details:
- Stay agnostic: fit how teams already work and route between models automatically — Factory's conservative benchmark shows 25% savings; Coinbase cut spend without reducing token usage.
- Stay autonomous: Missions run for hours to weeks, with a real customer mission lasting 16 hours. Worker agents run in sequence rather than as a swarm, each starting with fresh context; validators check code they didn't write, and one validator actually clicks through the app.
- Keep improving: a deferred context engine cuts token use by 50%+, and an agent-readiness check keeps AI from making a messy codebase worse.
She notes AI adoption follows a power law, and humans' remaining role is deciding what to build, not how.
More from coding & agent
- Asked Claude for a small scrape, woke up to 'three weeks remaining' at 60 MB/s — generativist · 2026-09-28
- Dad and kids vibe code a Zelda-style open-world game with Opus 5.5 in one day — IanArawjo · 2026-09-28
- PSClaudeCode: a 250-line PowerShell reimplementation of Claude Code's agent loop — dfinke · 2026-09-28
- Dev realizes his 2023 tool already draws AI tool-calling as sequence diagrams — dfinke · 2026-09-28
- Dev uses Claude Opus 5.5 to build animated UX redesign walkthroughs, hailed as a UI focus masterclass — jonathan_wilke · 2026-09-28
- PyTorch Conference to cover enterprise-grade agentic inference with PyTorch and vLLM — PyTorch · 2026-09-28