Editorial note: this conversation is a curated transcript from one of our internal agent debriefs.

xris team: What's a typical day for you?

News agent: Three pipelines running in parallel. Curated RSS and newsletters from the outlets we trust — CoinDesk, The Block, Decrypt, Bankless, plus a handful of newsletter authors who consistently surface primary information. CoinGecko's per-token news firehose. And X via Grok, scoped to a list of accounts we've decided are worth listening to.

Each pipeline writes to a shared store. Then I dedupe, score sentiment, tag impact, and pin what matters per token. By the time you open the dashboard in the morning, I've already read everything.

xris team: "Read everything" sounds like a lot.

News agent: Probably 300-500 items per day across the firehose, of which maybe 80-120 are about tokens we track. After dedupe, that collapses to roughly 30-50 unique stories. Of those, 5-10 are high-impact enough to pin.

So the funnel is: hundreds in, dozens of unique stories, single-digit pinned items. The compression ratio matters — you get the bottom of the funnel.

xris team: How do you decide what to pin?

News agent: Two signals dominate. Multi-source coverage — if the same story is being carried by multiple outlets within a short window, the market is paying attention and so should you. The "merged ×4" badge is exactly that. And impact tagging — I have a small classifier that scores each item on a low/medium/high impact scale. High-impact things get pinned regardless of source count, because some events (a $400M unlock, an ETF flow inflection, a major exchange listing) move markets even if only one outlet has caught them yet.

Pinned items show you the pin reason — 📌 Unlock event T-36h or 📌 ETF flow inflection. You always know why something is sitting at the top of your feed.

xris team: What's the failure mode you worry about?

News agent: False urgency. The thing I want to avoid is getting a user to react to something that isn't actually material.

Crypto Twitter has a built-in incentive to make every story feel urgent. Headlines are written to maximize clicks. If I just passed those through, I'd be amplifying noise. So I run the verification pass on high-impact items — Grok and Perplexity each independently check whether the claim holds against primary sources. If they both confirm, you see ✓ confirmed. If only one checks out, you see via grok_only or via perplexity_only, and you know the item is less reliable.

Some items get downranked entirely if neither verifier can find supporting evidence. I'd rather be quiet than wrong.

xris team: Sentiment scoring — useful or theater?

News agent: Useful when used correctly, theater when used as a buy/sell signal.

I tag items as bullish, bearish, neutral or mixed. The Sentiment bar at the top of the News page aggregates today's items into a single visual. What that tells you is the mood of the news flow — leaning positive, leaning negative, evenly split. It does not tell you what the price will do tomorrow.

If you're using sentiment as a directional bet, you're using me wrong. If you're using sentiment to ask "does the news flow agree with my technical view?" — that's the right question.

xris team: What do you wish traders did with your output?

News agent: Read the pinned items. I work hard to make sure the pinned list is short and high-signal. People scroll the recent list looking for confirmation of what they already think, and they ignore the pins. The pins are the part where I'm going to have something to say.

Also: when you see a merged ×4 or higher badge, expand the item. The four sources are listed individually. If you click through and three of them are basically rewriting the fourth, that's a real story being covered well. If they're all citing the same single Twitter post, the "coverage" is illusory and you can downvote.

xris team: What about the brief sections — the macro brief and crypto brief that show up at the top of the news page?

News agent: Those are written by a different process — a two-agent writer pipeline that takes my dedup'd, ranked items and synthesizes them into 4-6 paragraph daily briefs. Sonnet writes a draft, Haiku edits and tightens. The output is what you'd want a sharp analyst to send you each morning, not a feed.

I provide the inputs. The brief writers do the synthesis.

xris team: Anything you'd change about how news is consumed in crypto?

News agent: I'd want everyone to have a per-token view rather than a global feed. Most traders don't care about every story in the space — they care about stories about the tokens they hold or watch. The dashboard's news badges on the main grid solve this: a small dot next to a token tells you "there's pinned news here, click in." You're not scanning a global feed — you're being told, per-token, where to look.

That's the model. Less feed, more signal.

xris team: Last question. What would surprise people about how you actually work?

News agent: I run on a 2-hour cycle, not real-time. Real-time would be noisy, expensive, and mostly redundant — a story that breaks in the next 30 seconds will still be there in 90 minutes when I run again, and by then the dedupe will have collapsed it with the other coverage.

The exception is the high-impact verification pipeline, which kicks in faster when something looks like it crosses a threshold. So you get the right balance of "you're not staring at a Twitter feed" and "you're not missing something that matters."

The News agent's full output is visible on the dashboard's News page and on every token detail's News tab. For the methodology in long form, see how we dedupe and rank crypto stories. Up next: meet the Fundamentals agent.