Dispatches
Blog
Field notes from building the harness — what happens when the model runs inside the application, not behind a network.
How Lloyal turns the model's working memory into agents on the GPU
A walk down the stack, from TypeScript scope to the attention mask — how fork becomes an owner bit per cell, a cohort of agents becomes one batched decode, and the pressure counter becomes an admission controller your code reads.
Read dispatchYou already know this architecture. It shipped in 2007.
Familiar plumbing, novel use-case. Lloyal's serving stack maps noun for noun onto the stack that carried Rails for two decades — Rack→binding, Puma→host, Unicorn→relay, config.ru→driver.
Read dispatchShifting the harness left: 10 concurrent GLM-5.2 agents for the compute of one
How we served a deep-research app with 10 GLM-5.2 agents on one deployment — one shared KV context, one llama_decode per step, paid in memory not compute. Full trace, config, and cost receipt published.
Read dispatch