President & CEO @ycombinator —Founder @garryslist—Creator of GStack & GBrain—designer/engineer who helps founders—SF Dem accelerating the boom loop
San Francisco, CA
Joined January 2008
- Tweets76.4K
- Following6K
- Followers1.1M
- Likes254K
Pinned Tweet
Tech gave me everything I have
Its capacity to lift people into abundance is incredible and there is nothing like it
We must make that into prosperity for everyone
"I realized tech is this thing that can bring people out of whatever situation they're in and often into prosperity. And that's what I want for everyone."
@ycombinator’s @garrytan tells @emilychangtv how tech changed his family's life.
Watch here: trib.al/sxg1VGR
Keep Opus 5.5 as your coding default in Claude Code. On 52 coding tasks in Claude Code 2.1.322, Fable 5.2 solved about as many (82.7% vs 84.6%) but cost 2.80x as much per solved task.
Opus already solves most of these tasks, though, so telling the two apart will take harder ones.
github.com/garrytan/g-harnes…
Here's my open source harness eval repo and what it found
github.com/garrytan/g-harnes…
Y Combinator 总裁 Garry Tan 开源 Agent 外壳评测库 g-harness-evals，实测同一大模型换不同框架通过率差不多，但单次成本差了 2.5 倍以上
Y Combinator 总裁兼 CEO Garry Tan 在 GitHub 上开源了一个名为 g-harness-evals 的评测项目。大家平时选编程 Agent 时往往只盯着底层大模型，但真正把聊天模型变成干活 Agent 的，其实是模型外面的那层“外壳”（Harness）——也就是系统提示词、工具调用、记忆机制和权限控制。
Garry Tan 搞这个项目的逻辑很直接：把底层大模型（如 Claude Opus 5.5、GPT-6.1 Sol）和测试任务完全固定死，只换外面的 Agent 框架，看看到底哪家干活最好、花钱最少。目前被拉上实验台对比的主流框架有五个：Claude Code、Codex、OpenClaw、Hermes Agent 和 QM，同时还顺带测试了他自己做的两个外挂组件：知识库 gbrain 和技能包 gstack。
【首轮实测：成功率拉不开差距，账单差距巨大】
第一批公开任务跑下来，有几个非常反直觉的发现：
在 20 个开发测试任务上，7 组“模型+外壳”搭配的任务通过率全都挤在 71% 到 79% 之间。比如同样搭载 Opus 模型，Hermes 通过率是 78.8%，Claude Code 和 OpenClaw 都是 76.3%，QM 是 72.5%。Garry Tan 在报告里直接写明，这些分数的 95% 置信区间高度重叠，统计上根本分不出高下。
真正拉开差距的是钱。每成功解决一个任务，OpenClaw 平均要花 0.44 美元，Claude Code 是 0.22 美元，而 Codex 只要 0.17 美元。
为了搞清楚钱到底花在哪了，团队把各个框架连到一个虚假的模型接口上，一分钱不花地抓包了它们发出的第一个请求。结果发现，面对一模一样的任务，不同框架塞给模型的初始上下文体量天差地别：Codex 大约发了 12,900 个 Token，Claude Code 发了 18,416 个，而 Hermes 直接飙到了 20,613 个。而且在一次 Sonnet 模型的测试里，OpenClaw 账单里 87% 的钱都花在了提示词缓存（Prompt Caching）的读写上，真正模型输出只占 13%。
【塞进上下文的技能包，模型一次都没执行】
Garry Tan 还拿自己写的两个热门开源组件开刀，结果一点面子都没给自己留。
第一个是 gstack（一套模拟工程团队角色的技能文件）。测试发现，虽然 gstack 的技能说明文本在每次会话里都成功加载进了 Opus 5.5，但在 46 次测试里，模型实际执行技能步骤的次数是 0 次，反而让单次测试成本直接翻倍。团队试着加了一个拦截钩子（Hook），强制要求先跑失败复现才准改代码，结果 Opus 直接通过 Bash 命令写源码，绕过了钩子整整 40 次。只有在加上专门的启动钩子（Start Hook）后，Agent 才会乖乖调用启动命令。
第二个是 gbrain（团队共享知识库）。如果不专门按提问者的身份做权限隔离（Scoping），直接给机器人一个全公司通用的 Token，在检索时 OpenClaw 和 Hermes 分别吐出了 123 条和 121 条提问者根本无权查看的跨权限私密记录。只有把知识库严格限制在提问者自己的权限范围内，越权读取才会降到 0。此外，挂载 gbrain 时，只有 Claude Code 会把知识库自带的操作说明传给模型，其他几个框架全都直接丢弃了。
整个仓库目前做成了“无密钥可验证”的模式。每一项跑分、置信区间和账单明细都附带了完整的原始记录（Receipts），任何人都可以在本地用 Bun 和 Python 3 跑脚本复算核对，不需要消耗自己的 API Key。Garry Tan 也在文档开头坦诚披露了利益相关：QM 是 YC 孵化的项目，YC 投资了 Hermes，而 gstack 和 gbrain 是他自己的个人项目。
Jensen deserves due process, not whatever is happening right now. I signed.
#standwithjensen
Still hard to believe everything that's happened with Jensen.
After 33 years of building NVIDIA. I think he deserves a fair process, and we deserve the truth.
I signed this open letter to stand with Jensen. Hope you will too. 💚
standwithjensen.com
#StandWithJensen
Garry Tan retweeted
I'm still in shock, the CEO of Y Combinator open-sourced his agent brain on GitHub, 32.4K stars, one clone away
7 second brains you probably never noticed, all ready to clone right now:
1. gbrain · 32.4K stars
Garry Tan's personal agent brain. every fact tracks its origin, and every agent taps the same memory
github.com/garrytan/gbrain
2. Karpathy's llm wiki · 50K stars
the single-page idea file that kicked off this entire wave
gist.github.com/karpathy/442…
3. ai memory · 13.7K stars
leave Claude Code halfway through and resume inside Codex with nothing lost. memory sits in plain markdown
github.com/akitaonrails/ai-m…
4. COG second brain · 1.4K stars
35 skills and 10 agents, no database. just markdown files that reason
github.com/huytieu/COG-secon…
5. agent second brain
fire off a voice note through Telegram and it arrives as a linked entry in Obsidian, or a task inside Todoist
github.com/smixs/agent-secon…
6. open second brain
each night it converts the fixes you keep making into rules your agent obeys
github.com/itechmeat/open-se…
7. claude second brain
one npx command wires up Obsidian, Claude Code and local search in a single pass
github.com/jessepinkman9900/…
seven repos sitting in the open. your agent still loses you by morning
save this
Garry Tan retweeted
Your team wiki in Notion was up to date for about a week.
GBrain keeps it current for you: every call and email updates the right person's page, everything's linked, and when two facts conflict, it catches it.
gbrain.io/memory?utm_source=…
What if your own personal agent could recursively self improve as you use it? You can. I made the idea-minder skill to do exactly that.
gbrain.io/s/idea-miner-what-…
Garry Tan retweeted
how do you turn a repo with 1,000 stars into a product used by 1.2 million developers?
@firecrawl co-founder and CTO @nickscamara_ walked me through it.
two Brazilians in SF, one camera, and a lot to say about agents.
Garry Tan retweeted
I’m old enough to be the dad of half my YC batch. This is my second time around.
Here’s what I know now that I didn’t in W12.
Nobody knows what they’re doing. The first time, I was sure everyone else had it figured out. Their ideas sounded smarter. Their traction looked better. Then I watched batchmates pivot, rebuild and start over, just like we did. Now I know we’re all in the same boat, just trying to figure it out. Some people are better at hiding it.
An exit doesn’t give you the answers. Selling my first company to Dropbox made a few things easier. Investors take the meeting. People return your emails. But the product still has to work, and users still don’t care who you are.
Everyone says startups are hard. The first time, I nodded along and didn’t really believe it. Now I have the scars.
Consistency beats intensity. I’ve run every week for 200 weeks straight. Most of them I don’t feel like starting. Startups work the same way. You don’t win on the big days. You win by showing up every day.
Same nerves as 2012. Better shoes.
Garry Tan retweeted
the first thing that's breaking with so much code changes (our team is growing and we're increasing our output immensely with AI) is code reviews. Its cost doesn't make sense anymore.
We built our own agent with @agentrunai using @particle_api's particle-1 model. The results are great. ~ 80%+ drop in costs with better quality in our codebase. And I'm sure we can use our self-improving and workflows to get it's cost down even further.
Also, it's open source. You can host AgentRun yourself and make the code-review template yours. github.com/Parcha-ai/agentru…
Read more: teapawt.net/MiguelriosEN/status/21…
Garry Tan retweeted
Retell just crossed $100M in revenue run rate.
$10M to $100M in 17 months. Profitable. $4.6M raised.
From day one, we’ve focused on solving the hardest parts of deploying voice AI in the product itself, so customers can build, manage, and improve their own agents without depending on us.
Our bet: the future of the contact center belongs to the teams who run it.
Proud of our team. Grateful to every customer who believed in us.
Garry Tan retweeted
Introducing Summer Engine (YC F26). Summer makes it easy for everyone to make and play games together!
You ask for the game you want. Summer builds it, and you play it with your friends.
Last Christmas I tried to make a game with AI. The game looked great. But, then I wanted real multiplayer, analytics and monetization, and it became a huge pain! And don't get me started on publishing.
A year later it is still the same for most people. The games look much better, but you still get stuck with a prototype.
So we fixed it.
Today, tens of thousands of creators from all over the world have built games on Summer. The community has made more than 240,000 assets that every user can build with, making it one of the biggest free asset libraries in the world.
You own the game you make on Summer. You can sell it and publish it on Steam, Web, iOS, Android, and on Summer Games.
Making games should be as fun as playing them.
Try it at summerengine.com
PS. @garrytan made a game in a weekend about being a founder in the YC batch. You can try "FounderStar" on the link above.
Garry Tan retweeted
My conversation with Zach Reitano (@ZReitano), co-founder and CEO of Ro
In four years he took Ro from a $250M run rate to over $3B, with about a third fewer people
Zach rarely does long conversations like this, so it was fun to talk about the details of what that took and everything happening in healthcare
Plus, his answer to the closing question is amazing
We discuss:
- the state of the US healthcare system
- refounding Ro
- strong attract, strong repel culture
- the notion of jobs over lives
- the power of celebrities in marketing
- the importance of tailwinds (AI, GLP-1s)
- trying to recreate his dad in software
Enjoy!
TIMESTAMPS
0:34 How Much Should We Spend on Health
4:59 The Patient Is Not the Customer
10:24 Inside Ro's Business
18:28 The Future of GLP-1s
25:46 Jobs Over Lives
30:02 The Turnaround
47:43 Where Healthcare Goes Next
1:00:27 Recreating His Dad in Software
1:06:33 Marketing Into Stigma
1:19:06 Zach's Hottest Take
1:20:58 The Kindest Thing
