哪些任务不该交给 Agent,而应写成可重复的确定性脚本?

官方资料解读:路径能写死时,优先用确定性脚本、单次模型调用或预定义工作流,而不是让模型自己决定步骤。附适用/不适用与和既...

IvesFeng
2026年11月12日

Tencent Hy4 Hits the Charts Again with 10x Scaling: Before Agent Becomes the Top Open-Source Project, Let’s Clarify the Difference Between “Callable” and “Downloadable”

After Hy4 went live, the number of calls surged, the inference cluster expanded by about tenfold, and it made it onto Agent Arena’s top open-source re-citation list; however, the official weighting still states, “Coming soon...”.

IvesFeng
November 6, 2026

Should batch Agent tasks be immediately switched to a cheaper computer-use tier?

Comparison of GPT-6.1 Terra Public Pricing: Flagship Tier vs. Budget "Computer-Use" Tier in Terms of Completion Rate, Cost Per Task, Failure Rate...

IvesFeng
November 3, 2026

mem0 Real-World Test: After adding long-term memory to the Agent, there’s finally no need to re-establish the character’s backstory every time across sessions.

mem0ai/mem0 聚合评测:项目卡、适合/不适合、跨会话人设与偏好用例、何时换 Zep/自建向量;附 Quicks...

IvesFeng
November 2, 2026

Langfuse 实测:给 Agent 挂上 trace 后,我终于能定位是哪一步在烧钱

langfuse/langfuse 聚合评测:项目卡、适合/不适合、trace 与 cost 用例、何时换 Phoeni...

IvesFeng
November 1, 2026

When the agent messes up again, I check its permissions and tool boundaries first—I don’t immediately switch to a larger model.

Comparison of the "Escalation" model and "Tightening" the allowlist: After a failure, first write the permissions table, hook the subroutine, and replay three examples of privilege escalation; citing Anthropic / Clau...

IvesFeng
2026年10月31日

When an agent fails, it's usually not because the tools aren't sufficient, but because the scope of its actions wasn't clearly defined.

Comparison: Clearly Defined Boundaries vs. Heap Tools. Success rate does not equate to safe boundaries; write the envelope tonight and retest using out-of-bounds use cases.

IvesFeng
October 30, 2026

As the context window keeps getting bigger, I've actually started cutting down on the search segments.

Comparison: Stuffing vs. Fine-Sifting. A larger window doesn’t mean you should stuff it with more data; tonight, I’ll trim one version of the search fragments and keep the retest threshold.

IvesFeng
October 30, 2026

browser-use 实测:让本地浏览器替我点页面,哪些步骤稳、哪些必翻车

browser-use 开源浏览器 Agent:安装、公开页读 Stars 与文档摘命令实测,登录墙/验证码必翻车,和 ...

IvesFeng
2026年10月29日

Multi-agent systems seem exciting, but I’d rather start by clearly defining the scope of a single-agent tool first.

The "Three-Piece Set" for Aggregating Executable Boundaries: Tool Table, Confirmation Gate, and Prohibited Use Cases; can be submitted tonight—no empty talk about multiple agents.

IvesFeng
2026年10月29日

Long-context approaches are no substitute for structured knowledge bases—I’ll stick with RAG/folders when appropriate, and just feed in the full text otherwise.

Last Friday, I helped a classmate review the customer service knowledge base Q&A. He crammed all the support tickets from the past two years, the product manual, and the FAQs into a single 200K context window and asked, “Log in...”.

IvesFeng
October 28, 2026

Stop thinking of Skill, MCP, Project, and Custom GPT as the same thing—just choose based on what you need to do.

Last week, someone asked me another question that crammed three things into one: “Should I build a Custom GPT? Or write a Skill? Or start a project...

IvesFeng
October 28, 2026