notes on building with and around models, wherever the specific tech lands. pages are organized by concept; the projects these lessons came from — pub-search (search over atproto publications), phi (a bluesky agent), the slack bot in marvin's examples — appear as evidence within them.

contents #

prompt-caching — the prefix invariant, per-model floors, write/read economics, and measuring hit rate on anthropic and openai
retrieval/ — query-time mechanics: asymmetric embedding, rank fusion, synthesis between retrieval and prompt
memory/ — how agents keep state between runs: message archives, deliberate vs background writes, write-time curation
models/ — which frontier models run where, observed temperament, and what a swap changes; newest first
local-models/ — running models on your own hardware: serving, tool-calling, harness weight
cluster-the-2d-projection — reading structure off an embedded corpus: umap, hdbscan, work items from geometry
named-entity-recognition — what NER returns, why decoder confidence measures the label rather than the span, and what actually removes noise
reimplementing-inference — porting a model to another runtime: the feature extraction drifts, not the weights, and how to measure it
pi-extension-design — borrow established lifecycle semantics, implement only demonstrated needs, and turn each extension into harness knowledge
computer-use — what a driver can act on depends on what the app exposes; where the time goes; keeping the chooser bounded
agent-framework-stacks — how LangChain, LangGraph, Deep Agents, and LangSmith map onto Pydantic AI's split stack
terminal-agent-control — a terminal handle is an input channel, not a program: prompts can land in a shell after an agent exits

adjacent, filed elsewhere on purpose: protocols/MCP (MCP is a protocol first; it stays with atproto), storage/turbopuffer (vector-store operations are storage operations), and the agent-memory design notes that live in the bot repo's docs beside the code they describe.