Henry Quillin

Co-Founder @ (YC F25)
Check out

[Henry Quillin @H_quillin]

@0xMovez 8 months old tho

[Henry Quillin @H_quillin]

@jvheaney I got into YC with this idea (adding prompts to version control). Ended up pivoting away tho

[Henry Quillin @H_quillin]

@thekitze Aged like fine wine

[Henry Quillin @H_quillin]

@wholemars Why? Why would i ever use a model that’s not the best

[Henry Quillin @H_quillin]

@iyoushetwt R

[Henry Quillin @H_quillin]

@dkundel MCPs are terrible. That's why I built Sage as a cli
https://t.co/R3S4QqEboc

[Henry Quillin @H_quillin]

@umershafiqq @SolidRender Curious, what made you want to build for architects?

[Henry Quillin @H_quillin]

@yacineMTB I built sage for this exact reason
https://t.co/XwhJ4K1HAi

[Henry Quillin @H_quillin]

Just read an absolutely wild paper

Core idea: stop forcing LLMs to communicate through lossy, expensive text. Let them pass raw hidden states and KV caches directly (like telepathy).

All training-free, four simple pieces:
- Auto-regressive latent thoughts (predict next 4096-dim vector, not token)
- Full KV-cache sharing (lossless working memory transfer)
- Tiny projector (100 steps on random text) keeps generation stable
- Works in chains or trees

Results (Qwen, Llama, DeepSeek, etc.):
GSM8K: 90.1 → 96.6%
MATH: 68.4 → 77.2%
HumanEval+: 78.4 → 89.2%
GPQA Diamond: 52.1 → 57.4%
70–84% fewer tokens & 4.0–4.3× faster

Zero fine-tuning. Plug-and-play with Hugging Face or vLLM.

Thinking of applying this concept to the LLM council for coding agents I've been working on: https://t.co/fhvuzv7Rs4

Paper: https://t.co/ySCpYUZ4YT
Code: https://t.co/I2aOa9EUD8

[Henry Quillin @H_quillin]

@iyoushetwt This is exactly why I built Sage!
https://t.co/R3S4QqEboc

[Henry Quillin @H_quillin]

@hthieblot Got tired of Claude Code giving me AI slop, so I built Sage
https://t.co/R3S4QqEboc

[Henry Quillin @H_quillin]

@aftermagics Got tired of Claude Code giving me slop, so I built Sage
https://t.co/R3S4QqEboc

[Henry Quillin @H_quillin]

@jakobjelling Distribution is everything

[Henry Quillin @H_quillin]

@Niklas_Sikorra What about follows?

[Henry Quillin @H_quillin]

@donvito better but slower
use codex to review your claude slop with sage: https://t.co/R3S4QqEboc

[Henry Quillin @H_quillin]

@CodeByNZ hot take but 5.1 codex high >>>> opus 4.5

[Henry Quillin @H_quillin]

@weswinder Easy way to overcomplicate your codebase

[Henry Quillin @H_quillin]

I wanted a tool that was automatic, persistent, and separate from my main agent. That’s why I built Sage, which runs in a separate terminal and watches your coding agent in real time, automatically cross-checking every response with other models (currently just OpenAI models, Gemini & Grok coming soon).

Unlike MCP tools, Sage is a full-fledged coding agent. It reads your codebase, makes tool calls, searches the web, and remembers the entire conversation. Each review is part of the same thread, so it builds context over time.

Would love your honest feedback. Feel free to join our Discord to leave feedback and get updates on new projects/features https://t.co/tlQykKmMl7

[Henry Quillin @H_quillin]

Got tired of copy-pasting Claude’s responses into other models, so I built a tool that quietly critiques your coding agent's every move.
https://t.co/R3S4QqEboc
👇

[Henry Quillin @H_quillin]

So, I looked for ways to automate it. Tools like just-every/code replace Claude Code entirely, which wasn’t what I wanted.

I also experimented with having Claude call the Codex MCP after every response, but ran into a few issues:

- Codex only sees whatever limited context Claude sends it.
- Each call starts a new thread, so Codex has no memory of the repo or previous turns (can’t have a multi-turn discussion).
- Claude becomes blocked until Codex completes the review.

Other third-party MCP solutions seemed to have the same problems or were just LLM wrappers with no agentic capabilities.

Additionally, none of these tools allowed me to choose to apply or ignore the feedback, so it wouldn’t confuse the agent if unnecessary or incorrect.