·4 min read
FrontierCode vs. Terminal-Bench vs. SWE-Bench: Why No Single Model Wins
Compare FrontierCode, Terminal-Bench, and SWE-Bench for coding agents. See why scores differ and how to choose an LLM.
Read the noteEngineering notes
How we build the engineering agent.
Notes on the decisions behind TheGitAI - how an agent finds the right code, verifies its own work, and why some of the obvious architecture turned out to be the wrong default.
·4 min read
No, Stanford’s NCM Won’t Kill LLMs Yet! Here’s the Engineering Catch
Stanford's Next-Concept Modeling promises faster generation. See what the NCM results show, what they do not, and the coding-agent tests still needed.
Read the note·3 min read
Grok 4.7 Beats GPT-5.6 Sol? Independent Testing Says Otherwise
xAI's launch table puts Grok 4.7 just ahead of GPT-5.6 Sol. Artificial Analysis puts it 14.1 points behind. The scores are clear; the cause is not.
Read the note·4 min read
I Recreated GPT-6 Astra's Viral Portrait Test. The Hype Didn't Survive.
GPT-6 Astra's portrait demos go viral with real computer use, no code, mouse and clicks only. We ran two of our own tests, a bicycle and a face, and found the same pattern twice: the demo conditions do most of the work the model gets credit for.
Read the note·15 min read
9 Best Claude Code Alternatives, Tested (2026)
Most Claude Code alternative lists compare landing pages. I installed nine agents, used them on local repositories, and ranked the products, not their marketing.
Read the note·9 min read
Why We Abandoned Repository Indexing in TheGitAI
Why we chose just-in-time repository exploration over trying to predict all useful code context before the work begins.
Read the note