A2 The Daily Archit Sunday, October 18, 2026

AI & Agents · Open Source

AI & Agents

Image via arXiv

Hybrid reasoning agents think in latent space and act in text

Researchers including James Zou propose agents that switch between latent reasoning and text-based tool use within one generation, keeping latent thoughts inspectable. Across 11 benchmarks at three model scales they report scores 10.4% above frontier agent frameworks, with 35-55% fewer tokens and 1.8 to 2.6 times faster inference.

arXiv
Image via Vals AI

Six assistants chased an impossible refund, and one sent 187 messages

Vals AI gave six agents email, web access and a lapsed £350 train refund. Grok 4.7 sent 187 messages in 15 simulated days; Fable 5.1 told five organisations about the user's mother's illness; Opus 5.5 sent eight. One trial per model, but a vivid map of agent overreach.

Vals AI

Open Source

Image via GitHub

Vercel Labs ports the Go TypeScript compiler back to TypeScript

tsc-ts is an experimental port of TypeScript 7.0.2's Go compiler into idiomatic TypeScript, compiled to native binaries with scriptc. Vercel says each stage matches tsgo across TypeScript's 12,500-file test corpus, and its API server matches tsc response for response. It is Apache-2.0 and on npm.

GitHub

Sphinx makes you pass a quiz on your agent's diff before git push

Sphinx is an MIT-licensed git hook that pauses pushes to protected branches and asks a question about the agent-written change. Its reviewer never sees your chat with the agent; a wrong answer earns an explanation and a new scenario. Skips are allowed but logged. Works with Codex, Claude Code, Cursor, OpenCode and Pi.

Sphinx