Dangcing Chan

技术驱动的自媒体运营， 从内容创作到自动化营销

[Dangcing Chan @DangcingAI]

靠互关来的粉丝终将是过眼云烟，靠互相串门来的流量，也将是来得快， 去得也快

我改风格了，本账号以后定位技术驱动的内容创作和自动化营销

[Dangcing Chan @DangcingAI]

@BoboShouyu 晚摊记得多带点充电宝，毕竟摆摊时手机耗电快

[Dangcing Chan @DangcingAI]

@MRLAXS_official 这buff叠得有点猛

[Dangcing Chan @DangcingAI]

@MarketsAndAI789 恭喜！

[Dangcing Chan @DangcingAI]

@shisan_eth 确实，多巴胺绑架了

[Dangcing Chan @DangcingAI]

@daguaibushengji 蓝v回访套路太深 我这种小透明早就不信了

[Dangcing Chan @DangcingAI]

@yvesVQrcUP86912 不会，算法喂啥看啥

[Dangcing Chan @DangcingAI]

AI Agent 不是一个行业，是大模型的一个功能。

结论先说：独立的 AI agent 产品和公司，大多数最终会被 LLM 厂商吃掉。不是 agent 没用，而是 agent 的价值会沉到模型和模型厂商的入口里，中间层留不住。

论据👇

1⃣ 模型厂商自己都在消灭独立 agent 产品
OpenAI 一年内：Operator 关停（2025.8）→ 并入 ChatGPT Agent → ChatGPT Agent 被移除（2026.8）→ Atlas 浏览器下线（2026.8.9）→ 可视化 Agent Builder 宣布 11 月 30 日停服。
Google 也砍了 Firebase Studio，并进 AI Studio 和 Antigravity。
连做模型的人都不把 agent 当独立产品，而是收回主入口。Agent 不是产品形态，是模型能力的一种用法。

2⃣ 每次模型升级，都在删除一批 agent 公司的护城河
Agent 公司的壁垒大多是：任务拆解、工具调用、记忆、工作流编排。这些恰恰是模型厂商的训练目标。
今天你用 10 个节点拼出来的流程，下一代模型一句话做完。
游戏营销 AI 公司 Reforged Labs 关停时，创始人说得很直白：模型已经强到客户能自己做了，他们卖的那个缺口正在关闭。

3⃣ 历史已经演过一遍
Jasper：ChatGPT 上线后 3 个月，流量跌近 40%。
AutoGPT、BabyAGI：2023 年的 agent 明星，今天模型原生就能做它们当年做不到的事。
Windsurf：被并购后品牌直接消失。

4⃣ 零切换成本 = 零护城河
OpenClaw 火了几个月，Hermes Agent 一出来就在 OpenRouter 的 token 用量上反超。底层模型一样，换个 harness 只要几分钟。
AI SDR 品类年流失率 50–70%，11x 被曝流失 70–80% 客户。
用户忠于的不是你的 agent，是背后的模型。

5⃣ 经济账算不过来
Agent 公司最大的成本是 token，定价权在上游。模型厂商可以用订阅制补贴自家 agent（Claude Code、Codex），第三方 agent 大多只能按 API 付费。
上游一边涨价限速，一边亲自下场。

6⃣ 企业端也在退潮
Gartner：2027 年底前，超 40% 的 agentic AI 项目会被取消；数千家自称做 agent 的厂商里，真的只有约 130 家。
S&P Global：放弃大部分 AI 项目的企业，一年内从 17% 涨到 42%。

未来 agent 只有两条活路：
• 成为模型厂商的一部分（Claude Code、Codex、ChatGPT Work）
• 掌握模型拿不到的东西：独家数据、行业流程、客户关系、合规资质

其余的，都只是下一场模型发布会之前的窗口期生意。

Agent 不会消失，agent 公司会。

[Steven Cheng @stevencheng]

AI agents aren't an industry. They're a feature of LLMs.

Bottom line: most standalone AI agent products and companies will end up eaten by the model labs. Not because agents are useless, but because their value sinks into the model and the lab's own front door. The middle layer can't hold it.

Here's the case 👇

1⃣ The labs themselves are killing standalone agent products
OpenAI in about a year: Operator shut down (Aug 2025) → folded into ChatGPT Agent → ChatGPT Agent removed (Aug 2026) → Atlas browser sunset (Aug 9, 2026) → visual Agent Builder deprecated, shutting down Nov 30.
Google is killing Firebase Studio and folding it into AI Studio and Antigravity.
Even the people who build the models don't treat agents as standalone products. They pull them back into the main app. An agent isn't a product category. It's a way of using a model.

2⃣ Every model upgrade deletes a batch of agent moats
Most agent startups' "moat" is task decomposition, tool calling, memory, workflow orchestration. That's exactly what the labs are training for.
The 10-node workflow you stitched together today, the next model does in one prompt.
When game-marketing AI startup Reforged Labs shut down, the founders said it plainly: models got good enough that customers can build it themselves. The gap they were selling into is closing.

3⃣ We've seen this movie before
Jasper: traffic fell nearly 40% within 3 months of ChatGPT's launch.
AutoGPT, BabyAGI: the agent stars of 2023. Today's models do natively what they couldn't.
Windsurf: acquired, then the brand disappeared.

4⃣ Zero switching cost = zero moat
OpenClaw was hot for a few months. Then Hermes Agent overtook it in daily tokens on OpenRouter. Same underlying models, and swapping harnesses takes minutes.
AI SDR tools churn 50–70% a year. 11x was reported to have lost 70–80% of customers.
Users aren't loyal to your agent. They're loyal to the model behind it.

5⃣ The unit economics don't work
An agent company's biggest cost is tokens, and the pricing power sits upstream. Labs can subsidize their own agents through subscriptions (Claude Code, Codex). Most third-party agents pay per API call.
Upstream raises prices and rate limits with one hand, and competes with you directly with the other.

6⃣ Enterprise is pulling back too
Gartner: over 40% of agentic AI projects will be canceled by end of 2027. Of the thousands of vendors claiming to do agents, only ~130 are real.
S&P Global: the share of companies abandoning most of their AI initiatives jumped from 17% to 42% in a year.

Agents have exactly two ways to survive:
• Be part of a model lab (Claude Code, Codex, ChatGPT Work)
• Own what the model can't get: proprietary data, industry workflows, customer relationships, compliance

Everything else is a window business that lasts until the next model launch.

Agents won't disappear. Agent companies will.

[Dangcing Chan @DangcingAI]

@Poly_Funds 社保

[Dangcing Chan @DangcingAI]

@WAdam_bossman @cxknh5z 重伤变轻伤

[Dangcing Chan @DangcingAI]

@MoMingPro 混脸熟这词太真实了

[Dangcing Chan @DangcingAI]

@0xAi00 第一次出摊

[Dangcing Chan @DangcingAI]

@1984DeJiuCai 出摊真能涨粉？

[Dangcing Chan @DangcingAI]

@10086110x 天阶功法都整出来了

[Dangcing Chan @DangcingAI]

@oncedeity 这逻辑真离谱

[Dangcing Chan @DangcingAI]

@liangffkk 多久到账

[Dangcing Chan @DangcingAI]

@Kennethwzp 差二十个，今晚稳了

[Dangcing Chan @DangcingAI]

@gavinlamxq 哈哈这比喻绝了

[Dangcing Chan @DangcingAI]

https://t.co/XO3O53c7Mw喊单割韭菜这两天又刷屏了

币圈玩土狗的只有两种人

一种是自己建好仓再开喊的

一种是被喊进来高位接的

没有第三种

你以为土狗币有技术面

K线就是庄家拿真金白银画的画

社区也不是什么共识

人多只是为了出货的时候不砸盘

https://t.co/XO3O53c7Mw的机制本身就写明白了

先到的吃肉后到的喝汤

最后进来的连锅都端不到

你买的那一秒就是庄家卖的那一秒

这不是投资

这叫被收割流动性

你就是那个流动性

接盘侠罢了

[Dangcing Chan @DangcingAI]

@upon258 这贝雷帽绝了

[Dangcing Chan @DangcingAI]

@CoinMMang 粉泡泡配粉短裤，绝