NVIDIA's SoL-Pi pruning cost me 30% more
Ran it on real Hono bugs and every pruning rule lost to one boring habit: don't touch what the model already saw.
Every article
15 articles, newest first.
2026
Jev makes the small calls so the big model can cook
Typed answers in milliseconds for a fraction of a cent. Wired into five spots, kept only where it beat frontier models.
Leaderboards don't know your repo
How I'd test Claude Code, Codex, Cursor, OpenCode and Empryo on code you actually ship: fixes, the full bill, review time.
What even is an AI agent harness?
One rename, step by step, and every place the harness changes what you get.
I got tired of building panels, so I made the app morph
One JSON file, one new panel in the desktop app. You use it, the agent builds it 🙂
I let agents loose on my own app. They found 84 bugs
105 claimed, 84 survived two hostile reviewers. Plus a fixed suite that says the same thing every time.
One search cost $1.14, so I rebuilt the engine
86,000 tokens from a single search 😂 Rebuilt it and metered every dollar across six models.
My agent takes notes, and forgets the useless ones
Notes pinned to files. Open the file and they come back, ignore them and they archive. No vector database.
The Marionette: a cheap model pulls the strings
When it guessed causes, fixes dropped from 3/3 to 1/3. Now it only points at files: 3/3 and 11% cheaper.
My code map thought a formatter was my most important code
Static analysis was sure, git history disagreed. Now the Genome learns from how I work, and the formatter fell to #103 😅
Best AI coding agents in 2026, from a guy who builds one
Picked by the job you need done, with sources and dated benchmarks. Yes I'm biased, so check me.
pi said $6.25. Anthropic billed $9.19
Empryo said $7.06 and got billed $7.08. This is how I check every agent against the real bill.
Five real bugs, two agents, one bug won
Fresh fixes from hono, zod and ky that the models haven't seen. 7/10 vs 6/10 and one 8-minute hang.
2025
ast_edit: no more search and replace roulette
Edits by symbol name, so there's no search string to miss. Started as my Master's thesis.
The Genome: my agent knows your repo before you type
Every symbol, who calls it and what depends on it, ranked and in context from the first message.