Evidence · 17 entries · as text
What we know so far, with receipts.
Between basic AI use and the hype there is a dark patch where nobody shows the floor. This is what the show has seen there. Every line says where it came from: an episode and the second, or a source you can open.
Lessons
What the show has learned, one lesson per page, each resting on a moment in an episode.
A factory with no owner files tickets about itself
Both guests so far land on the same room: ownership. Vladimir keeps someone who owns the product and decides what not to build. Patrick keeps someone who owns the factory, because a team that also has features to ship will put the factory last.
Architecture decides what an agent can break
Assume the agent will find a way to delete the database. Limit its rights, cap how much it can change at once, and know how fast you can restore and how much you lose. Context and harness are not enough to survive production.
Make the first loop work before you automate it
A factory built on prompts and no checks does not remove the mistake, it repeats it at scale. Get the local loop reliable, then add verifiers and risk analysis, then optimise. Patrick calls the other order automating disaster.
Nobody has the right setup yet, including the people selling one
Said by a guest who builds and sells a factory. The whole industry is early. The way to get better is to run one and learn from the cases, not to wait for the finished pattern.
Structure agents like people
An agent left alone for two days produced code that did not work and tickets that did not matter. What fixed it was structure: dedicated agents with a job title and ownership of a task, like a team.
The bottleneck moves to the person who approves
When agents write most of the code, the wait does not disappear. It moves: first to review, then to CI, then to the one person who confirms each deploy. Plan for where the queue forms next, not for the queue you have.
When building is cheap, the decision is what not to build
A version takes five minutes. Vladimir builds what the business asked for, shows it, and throws it away if they do not want it. Patrick: the decision is more and more about which features not to build.
Factories
The factories our guests run or saw. The 75 we track across the industry are in factories.json.
Cobalt Code
Vladimir Petrov's factory. Started as a Lovable clone built for fun. The big clouds were too heavy to run it, so he built his own sandboxes: a container up in seconds, torn down, or a hundred of them. First real run: three months of open tickets, 25 pull requests at once, one server down.
The Tessl factories
Tessl ran a hackathon and ended up with five or six factories, to learn the shape of one. One took over the marketing and sales board. One morning the factory had filed 15 tickets about improving itself. Tessl then put a team on it, the way a platform team owns a cloud.
Tools
What guests actually use, with their words on it. Not a directory.
Harness
The code around the agent: the checks, the tools, the limits and the loop it runs in. Patrick's ladder for a company: prompt, context, harness, loop, factory. Context right and harness right is still not enough to survive production.
Sandboxes per agent
You do not share your PC with a teammate, so why share it with an agent. Cobalt gives each agent its own container, up in seconds and torn down after. Built because the big clouds were too heavy for it.
Skills registry
Shared skills behave like libraries: people fork a popular one, then face the choice of contributing upstream or maintaining their own. Tessl looks at which skills are similar, which are used, and what the agents' logs say about them.
The ticket board as the factory's intake
The factory reads its work from the board. At Tessl it read more than engineering: it took over the marketing and sales board in Linear and started commenting on the website. The intake defines the blast radius.
Verifiers and evals
Tests for what the agent did: linters, deterministic checks, adversarial tests. Tessl kept translating the notes in its context into checks the agent could not argue with. Add them right after the first loop works.
Events
Where this conversation happens.
AI Native DevCon 2026
Patrick Debois' conference. This year's theme is software factories and how to get there. Workshops on day one, then two days of talks. Patrick's discount code on the show: PATRICK50.
News
What moved in software factories. Two lines each, with the source.
EP. 03 with Milko Slavov lands on 29 October
Milko Slavov of felan.ai on which rooms to keep lit. He keeps three: what to build, the architecture, and the final sign-off.
EP. 02 with Patrick Debois is out
The second episode went live on YouTube and Spotify. Patrick on the factory that filed 15 tickets about itself, and why it cannot run without an owner.
How this grows
After each episode the pipeline drafts lessons, factories and tools from the transcript. An agent watches the show's channels and a list of sources for news and events. Everything lands as a pull request with a report, and a person reads it before it goes live. The format is in evidence.md and in the repository's wiki/EVIDENCE.md.