I will be publishing posts here
Horsies keeps the queue and the workflow engine on one transactional Postgres substrate: typed task contracts, first-class DAGs with joins and success policies, and a static validation gate that refuses to boot a broken graph.
We as humans can't read/write two functions at the same time, yet given the correct descriptions, agents can write tens or hundreds of them. And this is the part where you, as a software engineer, have to adapt.
The memory allocator is a library you've probably never configured, yet it can silently double your RAM footprint. Six allocators, one identical Python web app: the leanest settled at 1099 MB, the heaviest peaked at 2306 MB — while latency stayed similar across all.