新闻沙盒 · News Sandbox
每天自动推送最新 AI 进展
读到的，一键就能复现、试用、实验。
a·gent·ist /ˈeɪ·dʒənt·ɪst/ n. — one who commands AI agents to do real work; a super-individual whose ability is trained, not granted. This edition reads the frontier so you can become one.
Agentist · a daily on the agent frontier
2026 · 每日更新
AI 导师 · AI Mentor
AI 导师按你的目标带你动手造
卡住当场解决，名校知识体系为你定制。
训练营 · Boot Camp
有真人老师带的结构化训练营
第二期招生中 · 14 周 · 40 席。
AI 导师 · AI Mentor
跟 AI 导师，把 AI 亲手造出来
不是看课，是动手造。整套 AI 已经被理成一张图——你不用学完它，导师从中只挑你现在真正该走的那几步。
-
①
名校体系，理成一张图
961 个知识点、14 个领域，来自 Stanford、Berkeley、MIT 与资深工程师——成体系，不是零散教程堆出来的。
-
②
不用学完，只学你需要的那几步
告诉 AI 你的目标，它从图里只挑有限几步，每步都解决一个你此刻解决不了的问题。
-
③
边写边学，能实现 ≠ 知道
每一步都结合代码直接练。工程师最值钱的不是"知道"，是"能把它跑出来"。
-
④
系统化把关，不让你"自以为懂"
不预测不给跑；导师还会让你复述概念、当场考你——确保理解精确、不遗漏、不误解。
-
⑤
绝不卡壳，卡住当场解决
半途放弃，多半是一个难点当场解不了、越拖越散。AI 导师随时在，问题不过夜。
新闻沙盒 · News Sandbox
每天，前沿自己送上门——还能一键跑起来
不用你去找。全球最新的产业动态与前沿论文，落到前沿的那一刻就理好推给你——读不懂有 AI 带读，读到的环境已装好，一键就复现。
产业动态 · Industry News
397 条 · 按时间 · 折叠 ↑
-
Claude Opus 5 刷新 SWE-bench Pro 纪录
79.2% SWE-bench Pro、30.16% ARC-AGI-3，安全评测同步公开。
-
英伟达、微软、Meta 联手反对过度监管开源模型
开源阵营罕见统一发声，治理分歧浮出水面。
-
Alphabet 加码 AI 芯片与基础设施投入
领跑超大厂资本开支，微软、亚马逊承压跟进。
-
Claude 5 时代的上下文工程新规则
官方发布面向 Claude 5 的 prompt 设计与上下文组织指南。
论文 · Papers
arXiv 进展 · 已提炼 · 全部 →
-
SAT
RL -
Active
Vision
● 环境已自动装好，读完就能跑
① 每天推送
替你扫完全网
每天扫 2000+ 信息源，重要进展第二天就在你的头版，不用自己刷。
② AI 带读
读不懂就有人陪
悬停任意一条，AI 当场把这段进展讲透，接上你已有的背景。
③ 一键复现
读到的立刻能跑
环境已自动装好，真终端、真 agent，什么都不用装。
AGENTIST — published by PARALLIGHT · 每日更新于 agent 前沿
News Flow
Industry News
OSWorld 2.0 benchmarks computer-use agents on long-horizon workflows
A new real-world complexity bar for computer-use agents, exposing failure modes beyond GUI control. 108 long-horizon tasks, ~1.6h median human time, ~318 tool calls vs ~30 in v1.0. Best agent completes only 20.6% of tasks at 500 steps.
Dockerless: environment-free verifiers for coding agents
Cuts the setup cost of verifying agent code, making RL on coding agents cheaper to run. Removes the container dependency from the verifier loop; matches container-based verification quality at a fraction of the overhead.
Agent Index
Full index →Every new skill, MCP server, Claude Code feature, model and API — the moment it lands on the frontier, it lands here, and you can run it in our online IDE. A real terminal, a real agent, nothing to install. Reading is free; running is one click.
Claude Codelive
claude mcp add: plug Playwright, Notion, or a database into your agent
One command wires a new MCP server into the agent's toolbox — then use it live.
mcpRun in IDE →
Skilllive
zyte-web-data: give Claude Code web-scraping powers
A skill pack that teaches the agent to pull structured data off the web.
skillRun in IDE →
Modellive
Try claude-sonnet-5 on your own task
Switch models with /model in the online lab and run the new model live.
sonnet-5Run in IDE →
The Bench runs on Parallight's live sandboxes — new items arrive as the frontier moves.
The Leader · Editorial
What an Agentist is
An Agentist doesn't just use agents — they command them. They read a problem, aim an agent at it, verify what comes back, and ship what holds. The tools reset every week; the person who knows how to wield them does not.
The models will churn. Your ability will compound.
Read the full editorial → Past editorials →Published by Parallight · San Mateo
Search
2,000+engineers trained
13countries · 4 continents
14AI giants next door
1point · the Bay Area