From the team
Blog
Research, benchmarks, and field notes on making AI coding agents employable.
5 viral repos, distilled
A Go document converter that renders pages and reads them visually, an offensive-security skill library for Claude agents, an offline terminal AI-text detector, a deterministic agent-workflow graph, and a Swift/Metal runtime that runs 80B-parameter models on an iPhone.
Full teardowns, same rubric every day →
SAAGE: a deterministic agentic workflow engine, distilled
A PocketFlow-based agentic workflow engine with composable loop primitives, checkpoint/resume, and a git-ref remote lifecycle. Scored 62/100, the highest teardown so far.
Full teardowns, same rubric every day →
5 more viral repos, distilled
A Rust document parser spanning nine formats, a zero-dep SVG icon morpher, a Bluetooth phone finder, a Codex skill for editorial photo abstraction, and a benchmark harness for testing agents against offline SaaS replicas.
Full teardowns, same rubric every day →
5 more viral repos, distilled
A multi-tenant agent harness with an admitted sandbox gap, a voice coding agent that survives disconnects, a role-pinned delivery workflow shipped half-built, distillx's own plugin ecosystem turned inward, and a desktop app that turns recordings into agent skills.
Full teardowns, same rubric every day →
5 viral repos, distilled
Five repos trending on GitHub, run through a deterministic readiness scan and a five-persona engineering panel. Scores, gaps, and the one idea from each worth stealing.
Full teardowns, same rubric every day →
What coding benchmarks actually say under the hood
Read the SWE-bench family side by side and it stops being a scoreboard and starts reading like a manual for getting better results at lower cost. The catch: changing the model is the last and smallest lever.
Seven levers, cheapest first →
Routing down is harder than routing
A 400 on one of our own sessions, and what it taught us about downrouting between models: keep the body valid on the target, beat the cache you give up, and pick a safe moment to be cheap.
Three checks, not one →
Four hours, 8.4 million tokens in, 2.9 million compressed away. Nothing lost.
We tabulated one ordinary afternoon of prefex compression across four specialized compactors. Every lossy byte is recoverable. Here are the receipts.
34.3% effective compaction →