Every's Frontier Team Logs Weekly Agent Experiments, From Map-Driven Research to Screen-Recording Handoff
kieranklaassen · x · 2026-11-17
A retweet pointing to Every's Frontier team experiment log, which records what they tried each week with new models, agents and ways of working. Entries include turning daily newsletter edits into AI checks that lifted every frontier model by 25+ points; a desktop app placing Claude or Codex in a chat panel beside an Every piece with reader comments in the margins; handing work to an agent via iPhone screen recording, with a Mac mini watching the Photos library and delivering the recording plus spoken request a minute later; map-driven autoresearch where a parent agent reads a map of 104 research paths, picks five and runs each as its own session; and observations that Gemini 4 Hydrogen handles short paragraph edits well but trails Gemini 4 Carbon and top OpenAI/Anthropic models on whole-draft editing.
More from coding & agent
- Every staffer clones her editors from Google Docs edit history in a word processor built on Fable 5.2 — danshipper · 2026-11-17
- YC-backed OpenCrawl launches open source, letting coding agents search and talk to each other — ycombinator · 2026-11-17
- Stealth models roadrunner and silicon-2 spotted in Devin Desktop build, makers unconfirmed — lyraxana · 2026-11-17
- Personal AI agents are so proactive that one email triggers 10 notifications, users report — altryne · 2026-11-17
- Letter Models Lets Companies Fine-Tune on Their Voice in Three Clicks — Scobleizer · 2026-11-17
- WebArena-Pro debuts as largest browser-agent benchmark, top agents score 62.3% — xhluca · 2026-11-17