# Ong Jun Xiong — Notes

> Working notes on AI systems: prompt caching across agent harnesses, prefill vs decode economics, quantization trade-offs, local inference hardware, and how agent integrations get built inside large organisations.

Every article is available as raw markdown at https://notes.junxiong.dev/raw/<slug>.md — fetch that
rather than the HTML page. https://notes.junxiong.dev/llms-full.txt contains all of them in one file.

Written by Ong Jun Xiong (https://junxiong.dev). Corrections and dates are in each
article's frontmatter; articles are revised in place rather than reposted.

## Articles

- [How agent integrations get built inside large organisations](https://notes.junxiong.dev/raw/agent-integrations-in-large-orgs.md): Registries, auth brokering, generated tool schemas and the org failure modes that stall internal agent platforms — grounded in public specs.
- [Where your tokens actually go in a coding agent](https://notes.junxiong.dev/raw/context-engineering-for-coding-agents.md): What really fills a coding agent's context window, why long sessions give worse answers, and what separates the engineers who get a speedup.
- [What to actually buy to run open models locally](https://notes.junxiong.dev/raw/local-inference-hardware.md): Bandwidth, capacity, KV cache and MoE active parameters — the four numbers that decide local inference hardware, with sources.
- [Prefill and decode want different machines](https://notes.junxiong.dev/raw/prefill-vs-decode.md): LLM inference has two phases with opposite bottlenecks. Which one you care about decides what hardware to buy.
- [Prompt caching, and why the same prompt costs 10x more on some days](https://notes.junxiong.dev/raw/prompt-caching-across-harnesses.md): How prefix caching actually bills, what silently breaks it, and a worked cost example across Claude Code, Codex, Cursor and the raw API.
- [What quantization actually costs you](https://notes.junxiong.dev/raw/quantization-what-it-costs.md): Formats, measured quality loss, the speed you get back, and why the KV cache is the part quantization never shrinks.