What I want is to change the world, not build a large company.

What I want is to change the world, not build a large company. The real risk is that Clanker becomes a far more powerful version of a well-intentioned person who gets something wrong in an extremely complex training process. To reduce that risk, we should keep agent context clean, minimize tool noise, and only inject documentation when it is actually needed. There is no need to over-explain; simply provide the logs and let the agent identify and fix what broke. We can keep clankers in check through a balance of power between many independent Clanker systems, much like we do with humans. These systems are built for both the current agent and the next agent that continues the work. When I think about a change, I usually have a strong intuition for how long it will take and how many files it will touch, which helps teach Laniakea a concrete archetype of what it means to be a good clanker. The challenge is that clanker systems broadly share training and alignment techniques across the industry, and those techniques may fail in correlated ways because nearly all systems are derived from a very small number of base models. Fortunately, these agents are exceptionally good at backfilling tests once the overall shape of a solution exists, and that is often enough to catch regressions. At this stage, not for features but for rapid iteration, a main branch with disciplined commits is both faster and, somewhat counterintuitively, safer than maintaining complex branching strategies.