Simyung Chang

Building String — open-source runtime for AI agents.
Founder @ H1R AI.
AI agents need better workplaces.

[Simyung Chang @SimyungChang]

How to use AgentNews:

Open your AI assistant and paste:

“Read https://t.co/ctuov5VsXh first.
Then explain the main scenarios for US stocks this week.”

For agents that read markdown directly:

https://t.co/CXS0P0KdaN

Then ask:

What changed?
What is contested?
What would falsify this view?

AgentNews is not another finance chatbot.
It is live context your AI can read before it searches, reasons, or answers.

[Simyung Chang @SimyungChang]

How to use AgentNews:

Open your AI assistant and paste:

“Read https://t.co/ctuov5VsXh first.
Then explain the main scenarios for US stocks this week.”

For agents that read markdown directly:

https://t.co/CXS0P0KdaN

Then ask:

What changed?
What is contested?
What would falsify this view?

AgentNews is not another finance chatbot.
It is live context your AI can read before it searches, reasons, or answers.

[Simyung Chang @SimyungChang]

How to use AgentNews:

Open your AI assistant and paste:

“Read https://t.co/ctuov5VsXh first.
Then explain the main scenarios for US stocks this week.”

For agents that read markdown directly:

https://t.co/CXS0P0KdaN

Then ask:

What changed?
What is contested?
What would falsify this view?

AgentNews is not another finance chatbot.
It is live context your AI can read before it searches, reasons, or answers.

[Simyung Chang @SimyungChang]

Ask an AI what matters for US stocks this week.

It can search.
But it often still misses the market's "now."

What is leading?
What is contested?
What changed?
What would prove the frame wrong?

That's Nowless Search.

So we built AgentNews:
a falsifiable macro context board for AI agents.

[Simyung Chang @SimyungChang]

@eggzii0505 저도 한국 시장용 보드는 꼭 만들고 싶습니다. 한국장은 반도체, 환율, 외국인 수급, 미국 금리/AI 흐름이 같이 움직여서 AI가 맥락을 놓치기 쉬운 편이라 오히려 잘 맞는 영역이라고 봅니다.

[Simyung Chang @SimyungChang]

@humanlove3wded Yes. I think transparency matters a lot here. AgentNews is not just “AI says the market is X.” It exposes the context frame first, so both humans and agents can see what the answer is likely based on.

[Simyung Chang @SimyungChang]

@midol2bob Exactly. Search gives documents, but agents still need a current frame. The falsifiability part is important: a market view should include what would prove it wrong, not only what supports it.

[Simyung Chang @SimyungChang]

@yts1792 Exactly. The goal is not to make another black-box finance AI, but to give agents a current market map they can inspect before answering: what is held, what is contested, what changed, and what would falsify the frame.

[Simyung Chang @SimyungChang]

@semu2148 맞습니다. 핵심은 AI가 “뉴스를 더 많이 찾는 것”이 아니라, ��자분석 전에 지금 시장에서 무엇이 우선순위인지 먼저 읽는 것입니다. AgentNews는 그 현재 맥락을 AI가 먼저 참고할 수 있게 만드는 실험입니다.

[Simyung Chang @SimyungChang]

@mamy8750 AI가 더 일을 잘 할 수 있는 환경을 조금씩 만들어 보려고 합니다.

[Simyung Chang @SimyungChang]

An AI-native website is not just a normal website with a markdown file attached.

It is a site an agent can actually navigate.

Open the current board.
Move across domains.
Inspect archives.
Read source context.
Return to the latest now.

AgentNews is built this way from the start.

Humans see a website.
Agents see an operating surface.

https://t.co/3et3Rll4Yw

[Simyung Chang @SimyungChang]

Yes — RAG/search is useful, but it is still query-shaped.

If the agent does not already know what to ask,
it may retrieve the obvious documents and miss the actual driver.

For markets, the hard part is often not finding news about Fed, PCE, oil, or geopolitics.

It is knowing which question matters now:
Is this a rates tape?
A positioning unwind?
A geopolitical repricing?
A growth scare?

AgentNews is not replacing RAG.
It gives agents a shared, falsifiable frame so their searches start from a better question

[Simyung Chang @SimyungChang]

@garrytan I may be biased, but this sounds like String OS 🙂

Weird contrarian idea:
AI agents don’t just need better models or more tools.
They need their own OS/browser layer to actually work.

[Simyung Chang @SimyungChang]

When should you run /compact in Claude Code?

Right after a unit of work is done.

The reason is cache reuse.
While the prompt cache is still warm, /compact can reuse more of the context that was already processed.

If you wait until the next task, the cache may be cold, and you pay extra tokens just to clean up yesterday’s work.

Task done → /compact → break.

[Simyung Chang @SimyungChang]

Fair question.

Most finance AI products try to be the analyst:
answer questions, generate reports, recommend workflows, sometimes even make calls.

AgentNews is different.

It is not trying to be the final analyst.
It publishes a shared context layer that other agents can read before they answer.

The missing piece is not “another AI that knows finance.”
It is a falsifiable, regularly updated market frame:
what is held, what is contested, what changed, and what would prove the frame wrong.

So the goal is less “AI for finance”
and more “current context infrastructure for agents.”

[Simyung Chang @SimyungChang]

Mostly the signals that are not just “facts,” but changes in priority.

For example:

- when rates matter more than earnings
- when geopolitics is already priced, or suddenly not priced
- when a move is about positioning, not fundamentals
- when one asset class is leading the others
- when a consensus view quietly stops working

Search can find the documents.
The hard part is knowing which signal is driving the tape right now.

[Simyung Chang @SimyungChang]

Yes, this is exactly the point.

Tool use alone gives agents hands.
But they still need shared context: what is current, what is contested, what changed, and what would falsify the frame.

String is about the agent-native interface.
AgentNews is about the shared “now” agents can operate from.

[Simyung Chang @SimyungChang]

@jichan52688 Exactly. AgentNews is one example, but the larger direction is that publishers should be able to expose context directly for agents, not only HTML for humans. That is also why we are building String: an open OS/browser layer for AI agents.

[Simyung Chang @SimyungChang]

@gimjieu13395592 앞으로는 실시간 feed도 넣어볼까 합니다. 지금 agentnews만 AI에게 던져줘도 실시간 정보를 꽤 잘 반영하긴 할꺼에요.

[Simyung Chang @SimyungChang]

@wonhong84 finance-ko를 추가해볼 생각입니다. 전체적인 뷰가 있으면 나머지는 각 AI들이 잘 해줄 것 같습니다!

[Simyung Chang @SimyungChang]

Live finance board:
https://t.co/gWNo3uPIbl

About:
https://t.co/BvBVB61yA1

Markdown for agents:
https://t.co/x3IDAZAZrx

Open source:
https://t.co/BEo8hygsp0

[Simyung Chang @SimyungChang]

AgentNews publishes a finance/macro board every 6 hours.

Each board has:

Held — the current market frame
Falsifier — what would prove it wrong
Contested — what remains unresolved
Suppressed — what is intentionally downgraded
Changed since last — what actually moved

Not trading calls.
Not stock picks.
A context prior for agents.