THIS UC BERKELEY AND MICROSOFT PAPER IS F*CKING INCREDIBLE
This is a dev post classified by AGK Intelligence as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
THIS UC BERKELEY AND MICROSOFT PAPER IS F*CKING INCREDIBLE Microsoft and UC Berkeley just proved the agent trace control thesis from my post below Instead of burning tokens on retry loops, their black-box controller cuts coding agent costs by up to 30.9% with zero weight edits: most teams deploying coding agents like Claude Code, Codex, or Copilot CLI evaluate them on isolated tasks. in continuous production workflows, agents face recurring environment-specific costs that quietly drain thousands of dollars in API spend UC Berkeley and Microsoft Research analyzed 60 multi-turn trace runs an
Posted by marfin (5.5k followers) 50 min ago · 12 likes · 376 views · view the original post on X. Kept by the Dev Radar as AI dev tools.
More dev work like this
- 𝗛𝗲𝗿𝗺𝗲𝘀 Agent needs 8 GB RAM minimum to run on local model. 🤯 — @BkashJosi
- Building an internal AI app? You don’t have to start with a blank repo. — @DanKornas
- Qwen3.8-27B decodes code at 53.6 tok/s on one 16 GB RTX 5060 Ti and 154.5 tok/s on a… — @niklaslenz_ai
- DevX with NeoAstra is just phenomenal 🤩 using CodeAlta and GPT Luna 6 for a couple of… — @laurentkempe
- computer-repair-skill 这个开源项目很接地气，值得推荐。 — @bkdgiffug
- Your AI workflow doesn’t need five separate windows. — @DanKornas
- Claude Sonnet 5.5 vs GPT-6.1 Sol for computer use. — @coastyai
- TensorFold 1.2.0 decodes Qwen3.8-27B code at 268.6 tok/s on an RTX 5090, level with 1.1.2. — @niklaslenz_ai
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.2k posts from 5k X accounts over the last 21 days, 8.9k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-01 13:03 UTC. Full method.