An early look at a software factory design @0xthanh and I have been building to speed up the way we ship Engineers, tear it apart. What’s missing, what’s fragile, what would you do differently? Feedback here could help shape the future of @base. We will figure this out and we will accelerate (responsibly) 100x when we do

We don’t let agents one-shot production code, so why let them one-shot changes to your skills that could affect production code? Any custom skills you use should be version controlled and changes to them should go through the same rigorous review flow that the rest of your code does

Top learning from this week: If there are still findings after 4 agent review rounds on a PR, stop and re-scope. It means either the spec is too vague or the scope is too big

Growing up: figure out how to add 4.5 pages of fluff on top of my 2 paragraphs of thoughts so my essay met the page requirement Now: figure out how to remove 4.5 pages of AI fluff to reveal the 2 paragraphs of coherent thoughts School did not prepare me for this

My rules on unit testing heading into this week: - do not write tests that restate the implementation (getter/setter checks, “calls X with Y” mock assertions, snapshot tests, tests whose expected value is copied from running the code) - write a test only when it pins behavior that could plausibly regress: edge cases, error paths, boundary values, parsing/serialization, arithmetic, concurrency, bug fixes (one regression test per fix) - prefer one test exercising real behavior over several testing internals. Test through the public interface; avoid mocking anything you own - keep tests short. If setup exceeds the assertion, reconsider the test or the design - do not add tests to a change just to have tests. Zero tests is an acceptable answer for trivial changes; say so in the PR

Just found some skills in my setup that were having the agent handle things that could have been automated in a script Don’t spend tokens unnecessarily. Would recommend auditing your skills for this

Kind of want to start acquiring server racks at home to prep for the world where open weight models are on par with the frontier. Why can’t I just host my own ai without the token or cloud spend?

A better question: “how can I prevent the stupid stuff AI does 1-5% of the time?”

Makes a lot of sense. Scrap the unit tests! At what point are e2e tests mature enough to stop requiring human approvals if they aren’t changed or broken?

I still think clean code matters, but not for the same reasons it used to Humans no longer need to read and understand the code (or at least we won’t need to in the very near future) But less code is still better than more due to context bloat. Less context needed to accomplish a task gives us more efficient agent output!

Today marks 2 years at @base / @coinbase 😄 Learned a lot along the way and excited for what year 3 brings

Cobalt is coming