About This Blog
An author-led publication focused on AI agents, startup strategy, context engineering, and the operating systems around modern software work.
This blog is designed as a search-first, machine-readable publication layer inside tonylee.im. The goal is to publish analysis that compounds over time instead of disappearing into timeline-driven feeds.
Coverage centers on AI agents, developer tools, startup economics, AI industry shifts, and the practical workflows that shape product execution. Posts are written to be discoverable, linkable, and legible to both people and machines.
Related Posts
How I Pick AI Subscriptions by Budget and Role
Before you compare prices, decide whether you code for work. The same budget buys a completely different stack for builders and non-builders.
Kimi K3 Just Beat Fable 5 on Agent Benchmarks
Moonshot's open-weight giant is matching frontier agent scores at Sonnet 5 prices, and the weights land July 27. Here is what actually changed.
If You Want an Agent Transition, Spend Time on Initial Setup
Boris Cherny's note on automation matches what I have been telling teams for weeks: the leverage is encoding judgment into infrastructure, not burning tokens on the same bug forever.
Starting Codex? Open This Guide First
I keep a living Codex Vible for install, settings, App vs Web, and how it differs from ChatGPT, because the product ships updates faster than most people can relearn it.
What I Told Huddling Club About Two Months as a Solo Builder
Your own channel, a work harness, and documented taste matter more than raw token spend. Notes from a weekend talk with other builders.
Builders Are Already Patching Codex for Performance and Quota
Two structural issues in Codex defaults cut reasoning short and explode agent trees. Here is what I measured and the config.toml changes that fixed both.
I Mean This Bluntly: Humans Have Less Work Left in the Loop
Orchestrators absorb more execution; the scarce skill is directing systems.
Claude Code Was Billing Tens of Thousands of Hidden Tokens Per Turn
Matt Pocock's method: run /context, rank tool payloads, then trim settings.json with disable flags and bare tool-name denies so unused definitions stop riding every message.