#llm
Every note tagged #llm, newest first — or browse the full archive.
Four terminals and an inbox: growing agent skills from their own message history
I ran an app-development workflow as four separate Claude Code sessions — each owning one role, all talking through a shared inbox — and then pointed a fresh session at the accumulated messages and asked it to write the skills. This is how the bootstrap worked, why the distillation step is the interesting part, and what it cost.
Counting the inbox: what an agent message corpus records about roles nobody designed
An earlier note described distilling four agent skills out of a shared on-disk inbox, and admitted that the transcripts were neither counted nor quoted. This one counts them and quotes them — how long each code name lasted, how the traffic decayed once the skills existed, and what the sessions wrote to each other about promotion, protocol, and disagreement.
Temperature zero is not determinism: your logits depend on who else is in the batch
Enumerating every ordering of a signed sum shows that reordering moves the result by one ulp — far too little to explain why temperature-zero endpoints return different paragraphs. The rest of the story is batch invariance, bfloat16, and argmax.
Running my day job on Claude Code — an agent architecture for non-software work
My day job isn't writing software — it's coordinating insurance-restoration jobs, which is document-heavy and relentlessly repetitive. Over a few months I built a system of Claude Code agents around that workflow. This is the architecture, the decisions that made it hold together, and what I'd change.