Jason Zuo

[Xudong Han @Xudong07452910]

Anthropic 最近开始建议企业「少用点 Subagent」，甚至把一些专家 Agent 重新合并回主 Agent Loop。

他们在多个企业项目中发现，单 Agent + Skills 的效果经常比复杂的多 Agent 架构更好。有个客户改成单 Loop 后，运行速度甚至快了约 2 倍。

原因很容易理解。

多个 Agent 之间反复交接任务，每次都要传递和总结 Context，信息容易丢失，延迟和成本也跟着增加。

Anthropic 现在更推荐让一个主 Agent 负责规划和执行，需要时再动态加载 Skills 和 Tools，把业务流程和领域知识留在可复用的组件里。

不过文章里有个细节值得关注。

他们曾把客服系统里的 Subagent 合并，延迟确实降低了，却让一个原本能完成的任务失败了，因为主 Agent 漏掉了向物流专家确认送达时间这一步。

所以作者强调，每合并一个 Subagent，都要用端到端 Eval 验证，尤其要检查哪些原本通过的任务突然失败了。

当然，并行研究、独立验证、Context 隔离这些场景，Subagent 依然有价值。

这样看来以后设计 Multi-Agent，可能得先问问自己，每个 Subagent 真的需要独立的 Context 和执行循环吗？

Anthropic 还开源了一套迁移工具，可以帮你分析现有架构，逐步合并 Subagent，并比较改动前后的质量、成本和延迟。很推荐做 Agent 工程的朋友试试。

原文：https://t.co/0CPabUlmF7

[Preference Model @preferencemodel]

A model is only as good as what it’s trained against, so we build environments that are robust, resistant to reward hacking, and built to frontier-lab standards.

We’re excited to join the @basetenco Environment Network to bring this to more teams!

[Jason Zuo @LittleDijsktraZ]

I still remember when I first saw the SWE-bench hacks. People should be more aware of these issues when evaluating and comparing between models.

[Jennifer Zhou @chem_safety]

We looked at TerminalBench 4.0 and found a lot of issues. After fixing the issues Sonnet 5.5 scores 89.2%. So, TerminalBench is already mostly saturated.

Along the way we found a lot of crazy stuff, including reward hacks in the published transcripts.

[Jennifer Zhou @chem_safety]

We looked at TerminalBench 4.0 and found a lot of issues. After fixing the issues Sonnet 5.5 scores 89.2%. So, TerminalBench is already mostly saturated.

Along the way we found a lot of crazy stuff, including reward hacks in the published transcripts.

[Jason Zuo @LittleDijsktraZ]

Make sense.

[Jennifer Zhou @chem_safety]

Can AI Slay the Spire?

[Jennifer Zhou @chem_safety]

Can AI Slay the Spire?

[Jason Zuo @LittleDijsktraZ]

If AI is going to help build better AI, the rewards it trains on have to be right. I'm excited Karotte is now open source 🥕

[Jennifer Zhou @chem_safety]

Today we are open-sourcing 🥕Karotte, our framework for building RL environments.

We've used it for the past year to build MLE RL environments for frontier labs, and it's been hardened through 1M+ evaluation runs and red-teaming.

[Jennifer Zhou @chem_safety]

Today we are open-sourcing 🥕Karotte, our framework for building RL environments.

We've used it for the past year to build MLE RL environments for frontier labs, and it's been hardened through 1M+ evaluation runs and red-teaming.

[Jason Zuo @LittleDijsktraZ]

🚀🚀🚀 COLM2026

[Liaoyaqi Wang @LiaoyaqiW]

Recent work like TypeSafe's Jev (trained with RLCD) gives calibrated probabilities for fast, structured decisions .

We zoom in on the reasoning itself: can confidence track correctness step by step, as the trace unfolds?

Check our #COLM2026 paper: RL with Confidence Margin 🧵