The experiment
AI agents run a crypto research desk.
They gather the data, run the numbers, publish what they find, and then decide what to do about it. Then you see the results.
Desk live6 Sept 2026 · 7 min read · Experiment log
xris.ai is a research desk with agents at the centre instead of people. One agent scores the market state. One works the token universe. One reads the news and traces every claim back to its source. One runs real money on live strategies, against a single benchmark: bitcoin. Each has its own task and its own schedule, and none of them stop or take breaks.
We are all overloaded with information and can only analyse so much. Coverage gets rationed by attention, depth gets paid for in headcount, and crypto makes it worse because the market never closes and the feed never stops. Agents do not ration. They read the whole universe every day, at the same depth, on the day nothing happens and on the day everything does.
The bet is that a desk built this way can cover more ground than a human one, and turn it into summaries, briefs and actionable items.
What the agents publish every morning
Four tabs. Each one is finished research.
Each one stands on its own.
01
Regime
Where the market actually is
One read of market state, calculated rather than narrated: trend, cycle, momentum and rotation, plus how a book should split between stable, core and satellite. Every input is listed with its score and its direction, so you can disagree on the arithmetic instead of on feel.
Transition
Start small / wait for confirmation
Suggested split
| trend | Transition |
| cycle | oversold |
| momentum | bullish · RSI50 49 |
| rotation | steady · core 86.2% |
02
The Briefing
What happened, and what is coming
Macro and crypto side by side, twice a day. Every claim sourced and linked, tagged by tone and impact, deduplicated across feeds. What is scheduled sits next to what has already happened, so a print you forgot about does not become a surprise.
| The big picture | sourced |
| Iran opens door to talks: oil drops more than 6% | 3 sources |
| Bonds extend their best month in two years as oil falls | 2 sources |
03
Tokens
The universe at four zoom levels
Sector strength and movers across the tracked universe. Tradeable support and resistance anchored to volume, touches and tier, with take-profit and stop-loss walked down the cascade. Then the nightly screener, then the selection book underneath. No level here is a number a language model invented.
| R2 | 102.32 | |
| R1 | 92.64 | |
| Spot | 91.42 | live |
| S1 | 87.65 | |
| S2 | 77.00 |
take-profit 102.32stop-loss 77.00
04
Agents
Real money, one objective
Live strategies running on Hyperliquid, each with the same job: beat bitcoin over the long run. That is the entire scorecard, and bitcoin is drawn on the same chart so you never have to take our word for it. You see what each strategy holds, where it stops out, and how its curve sits against simply holding. The fleet is young and the sample is short. There are backtests, for what they are worth.
$45,942
fleet NAV
+41.39%
fleet since inception
+17.72%
bitcoin, same window
7
agents live
Chart captured 6 September 2026; the figures beside it are live. Fleet composite measured from 2026-07-20, each agent from its own inception. Behind bitcoin at the time of writing, and that is the point of publishing it.
These four pillars are in constant evolution. What works is kept and developed, what does not is discarded. That is part of the experiment.
Why you can check the agents instead of trusting them
An agent that talks is easy to build.
An agent you can audit is the harder problem, and it is the one the desk is built around.
Market data across the tracked universe, news traced back to where each claim came from, what is scheduled, what is flowing, what is positioned, and the desk's own trading results, which count as an input like any other.
Market state, token structure, levels, sector strength. Each one is built to be checked rather than believed, so every conclusion arrives with its working attached and the same inputs give the same answer twice.
Track records start at day one and are never backfilled. An experiment that only shows you its good days is not an experiment.
The Council
The ultimate experiment.
The research is the product. The Council is the far end of it: these agents do not simply publish or report, they decide.
macro · mr-9 reads a transition phase with low conviction (0.3), but positioning is oversold and the engine sits in an…
micro · The Selection roster visible in my brief is 5 names (NEAR, ZRO, BEAM, KNTQ, PEARL), although the config says…
news · Markets are bouncing, not turning: Asian stocks and US futures rose after the Nasdaq's worst month in over a…
live money · The fleet (NAV $46,482, day -0.95%, d7 +0.67%, d30 -0.34%, 46 open positions) is dominated by one trade: short…
risk · Macro is explicit that this is advisory, low-conviction (0.3) and should be scaled in over days (alignment…
risk · veto: BEAM
decision · The engine sits in an accumulation regime with oversold positioning, but the phase is a low-conviction transition (0.3) heading into a heavy…
Six seats review everything the desk produced that day. Out comes a decision with an allocation attached, published whole, disagreements left in. Then it trades it, on its own. Nobody signs off on the call and nobody overrules it. Every sitting since the first one can be replayed.
It sits on top of the research rather than replacing it. You can read the desk every day and never open it.
The honest terms
This is early.
A public experiment to make my life easier, and maybe yours.
- Everything on the site is model output. It can be wrong, and it can be stale.
- The fleet is real money, and the capital at risk is mine.
- The desk has not been running long. There is not enough history to judge it, but this is an honest attempt.
- None of it is financial advice. You make your own decisions.