Discover and learn

09 Nov 2026 • Jack Rudenko

Your AI agent can pass every health check and still confirm a seat nobody checked. What Google's AQuA shows about AI agent observability, and how to turn production failures into regression tests.

05 Oct 2026 • Jack Rudenko

A 4B model beat both 9B browser agents we tested. We benchmarked 17 local vision models on screen understanding, UI grounding and OCR, and verification mattered more than size.

27 Sep 2026 • Jack Rudenko

Anthropic published per-task effort data for Terminal-Bench 3.0. It matches what we saw in our own evals: more effort helps on average and hurts on specific tasks.

24 Sep 2026 • Jack Rudenko

Jev came out of stealth on 15 September and every demo says the same thing: look how fast, look how cheap. Both halves are true. And most of those demos picked the wrong opponent.

11 Sep 2026 • Jack Rudenko

You give the agent ten steps. It does two and stops. We deprecated Magus's first two plugins — GTD and Kanban — because the models grew up. The board went back to people.

12 Aug 2026 • Jack Rudenko

A few months ago I called long-running AI agents a gimmick and a fast way to get hallucinated code. Now I keep two or three of them running for days. This is what changed.

27 Jul 2026 • Jack Rudenko

New AI models ship almost every day, and keeping agent tools up to date became a second job. So we built models.madappgang.com — a model catalog server for humans and agents, updated twice a day, human supervised.

16 Jul 2026 • Jack Rudenko

A regex query classifier that routes symbol lookups away from LLM expansion recovered 4x retrieval quality in MadAppGang's RAG pipeline ablation study. Here's the architecture, the data, and the decision tree for production use.

15 Jul 2026 • Jack Rudenko

We ran 860 queries across 12 repositories to test LLM query expansion in code search. Every LLM component hurt quality. One regex rule helped. Here's the data.