The Brief

New briefs daily around 7 AM Eastern

The score is not the agent: the day's gains sit in harness design, not size

Meta's return to open weights is the loudest release of the day, but the number that matters is the gap. Muse Glimmer posts a competitive intelligence score while trailing badly on agentic evaluation and hallucination, which is exactly where an agent loop lives. That split runs through the rest of the day: the results that actually moved were harness results, from a coding agent that rewrites its own core under review, to a document environment that treats long-form reading as mutable state and pulls small models toward frontier performance, to teams deliberately deleting most of their prompt and tool surface. A cluster of papers argues the instruments are the weak link, finding that passing tests are a poor proxy for developer intent, that verbalized confidence misses errors internal probes catch, and that static hallucination benchmarks overstate robustness once you fuzz them.

• 01Release

  Meta returned to open weights with Muse Glimmer, an Apache-2.0 30B that reportedly matches a 1T-parameter peer and runs on a single GPU at full context.

  Artificial Analysis→simonwillison.net→Bijan Bowen→
• 02Watch

  The headline index hides where that model is weakest. Its gaps concentrate in agentic evaluation and hallucination, so a 953 Elo intelligence score is a poor reason on its own to swap it into an agent loop.

  Artificial Analysis→
• 03Method

  Harness design is doing the work model scale used to. Ouroboros evolves its own core under review, DocAtlas reframes long-document reading as mutable-state interaction and trains small models near frontier results, and Databricks proposes Omnigent as an Apache-2.0 meta-harness to combine the rest.

  Anton Razzhigaev, Andrei Gritsaev, Andrei Kaznacheev, Nikita Dragunov, Roman Yampolskiy, Andrei Kuznetsov→Hongchen Wei, Yuanzhe Wang, Bei Liu, Yifan Yang, Qi Dai, Kai Qiu, Yunsheng Li, Dongdong Chen, Chong Luo, Zhenzhong Chen, Baining Guo→databricks.com→
• 04Method

  Prompt and tool surfaces are shrinking on purpose. The Claude Code team cut 80% of its system prompt, Nous replaced twelve browser tools with a single code-writing one, and a widely shared experiment got better results by writing the agent a plain-language letter.

  Y Combinator→Nous Research→Theo - t3․gg→
• 05Debate

  Verification is the day's real bottleneck. DevIntent quantifies how often passing code still violates what the developer meant, Tangent maps what agent test suites miss, probes catch errors verbalized confidence does not, and fuzzing shows static hallucination benchmarks overstate robustness.

  Susana Haing, Natan Vidra, Spurthi Setty→Rangeet Pan, Tyler Stennett, Divya Sankar, Bridget McGinn, Alessandro Orso, Raju Pavuluri, Saurabh Sinha, Maja Vukovic→Jyotin Goel, Ipshita Bandyopadhyay, Justin Shenk→Pengfei Zhou, Jiajun Song, Zhiwei Tang, Yixing Ma, Xiaopeng Peng, Donghui Si, Yuhang Xu, Huiqi Song, Yiyuan Miao, Yichen Qian, Weihua Chen, Wangbo Zhao, Bohan Zhuang, Jiasheng Tang, Yang You→
• 06Tooling

  Trust boundaries moved in two directions at once. Claude Code made auto mode the default execution posture, while OpenAI restricted its new frontier cyber model to approved defenders.

  ClaudeDevs→OpenAI→
• 07Tooling

  Cost and throughput housekeeping worth banking: Sonnet 5's introductory pricing is now permanent, RL rollout fleets can leave the trainer's cluster because under 1% of served weights change between versions, and host OS choice is a 3x lever on shell latency inside the agent loop.

  Claude→AI Engineer→SemiAnalysis→