What It Actually Takes to Build a Software Factory — Tereza Tížková, Factory
AI Engineer
93,157 views
Everyone is talking about software factories, but few people are building one. Tereza Tížková of Factory, which runs software factories for enterprises like EY and Adobe, defines one as the whole software lifecycle run autonomously: collecting signals, prioritizing, building, validating and improving. It's not just a swarm of coding agents, because writing code is the easy part.
She lays out three principles. Stay agnostic: fit how your team already works, and route between models automatically (Factory's conservative benchmark shows about 25% savings). Stay autonomous: Factory's Missions run for hours or even weeks, with a real customer example running 16. Worker agents run in sequence instead of a swarm, so each starts with fresh context, and validators check code they didn't write. One of those validators actually clicks through the app. Keep improving: a deferred context engine cuts token use by 50% or more, and an agent-readiness check keeps AI from making a messy codebase worse. She closes on what's left for humans: deciding what to build, not how.
Speaker info:
X/Twitter: @tereza_tizkova (https://x.com/tereza_tizkova)
Website: https://www.terezatizkova.com
Related links:
Factory: https://factory.ai
Timestamps:
0:00 Intro
0:45 What we'll cover
1:25 What a software factory is
2:05 Why it wasn't possible before
2:55 What a software factory is not
3:50 Three principles
5:10 Agnostic: fit how teams already work
5:45 Coinbase: lower spend without fewer tokens
6:45 Automatic model routing
7:35 About 25% savings
7:50 How the router works
8:40 Common routing questions
9:20 Caching
10:10 Autonomy and loops
10:40 Defining "done"
11:20 Long runs and cheating agents
12:05 Factory Missions
12:35 A real 16-hour mission
13:00 Workers in sequence, not swarms
13:40 Validators and the validation contract
14:05 A validator that clicks through the app
15:20 Always improving: the context problem
16:35 Deferred context: 50%+ fewer tokens
17:25 AI adoption is a power law
18:20 Agent readiness
19:20 Plugins and unwritten team knowledge
20:20 What happens to humans
21:30 Let agents take the annoying work
20 Comments
It feels SO refreshing and building everything manual again 😍😍
What It Actually Takes to Build a Software Factory — Tereza Tížková, Factory
AI Engineer
93,157 views
Everyone is talking about software factories, but few people are building one. Tereza Tížková of Factory, which runs software factories for enterprises like EY and Adobe, defines one as the whole software lifecycle run autonomously: collecting signals, prioritizing, building, validating and improving. It's not just a swarm of coding agents, because writing code is the easy part.
She lays out three principles. Stay agnostic: fit how your team already works, and route between models automatically (Factory's conservative benchmark shows about 25% savings). Stay autonomous: Factory's Missions run for hours or even weeks, with a real customer example running 16. Worker agents run in sequence instead of a swarm, so each starts with fresh context, and validators check code they didn't write. One of those validators actually clicks through the app. Keep improving: a deferred context engine cuts token use by 50% or more, and an agent-readiness check keeps AI from making a messy codebase worse. She closes on what's left for humans: deciding what to build, not how.
Speaker info:
X/Twitter: @tereza_tizkova (https://x.com/tereza_tizkova)
Website: https://www.terezatizkova.com
Related links:
Factory: https://factory.ai
Timestamps:
0:00 Intro
0:45 What we'll cover
1:25 What a software factory is
2:05 Why it wasn't possible before
2:55 What a software factory is not
3:50 Three principles
5:10 Agnostic: fit how teams already work
5:45 Coinbase: lower spend without fewer tokens
6:45 Automatic model routing
7:35 About 25% savings
7:50 How the router works
8:40 Common routing questions
9:20 Caching
10:10 Autonomy and loops
10:40 Defining "done"
11:20 Long runs and cheating agents
12:05 Factory Missions
12:35 A real 16-hour mission
13:00 Workers in sequence, not swarms
13:40 Validators and the validation contract
14:05 A validator that clicks through the app
15:20 Always improving: the context problem
16:35 Deferred context: 50%+ fewer tokens
17:25 AI adoption is a power law
18:20 Agent readiness
19:20 Plugins and unwritten team knowledge
20:20 What happens to humans
21:30 Let agents take the annoying work