# everyone needs a quant. now any agent can ask one By [tomu](https://tomu.xyz) ยท 2026-07-09 --- markets eventually price everything they can measure. price came first, then volume, then order flow, then onchain data. every time one of these became easy to read, the edge inside it got competed away. that's most of quant finance in a sentence: find the data nobody has structured yet, structure it, trade it, and watch everyone copy you until the edge is gone. right now, the last unpriced dataset in crypto is attention. which sounds strange, because attention is all ct talks about. narratives, mindshare, the timeline turning on something. but talking about attention and measuring it are not the same thing. what people call sentiment analysis today is really just counting, mentions per hour, positive words, likes weighted by follower count. counting attention is like judging a market by how many trades happened while ignoring who was buying, who was selling, and how big. the number goes up and tells you nothing about what comes next. here's the idea everything i've built rests on: attention has a shape, that shape can be measured, and once you measure it, it moves before price does. not as a metaphor. by hours. **a year of watching the crowd** -------------------------------- i've been running [checkr](https://x.com/checkrsocial) for over a year. quietly measuring the attention around every token it tracks. the tool underneath it is a hawkes process, a model for figuring out whether one event makes the next one more likely. it's the same idea used to study aftershocks after an earthquake, or one trade setting off the next in a fast market. point it at social posts and it answers a simple question: is this crowd actually feeding on itself, or is it just noise dressed up to look like a crowd. that question is the whole edge. anyone can buy volume. a bot farm will post as much as you pay it to. what it can't fake cheaply is the rhythm of real people reacting to each other in real time. that rhythm is what checkr reads. volume lies. the shape underneath it mostly doesn't. so checkr doesn't count posts, it reads them. and what comes out is the stuff that actually matters. whether a wave of attention is real or manufactured. whether it's carried by a whole crowd or by one big account that can kill it the moment it goes quiet. how much bigger the wave gets before it dies, and roughly when. and the pattern that pays best of all, attention climbing while price hasn't moved yet. that's the thing a year of data keeps proving. attention leads price by two to six hours. the crowd shows up first, the candle follows. the signal is sitting right there on the timeline, in public, hours before it reaches the chart. everyone can see the posts. almost nobody is doing the math on them. but two to six hours is the average, and the average is shrinking. as more capital learns to watch attention, the gap between the crowd forming and the price moving gets tighter. **why that edge is useless to you** ----------------------------------- a two-hour head start sounds like plenty. it isn't, not once you try to use it by hand. you spot the signal somewhere in the middle of the window, if you're lucky, after it's closed if you're not. now the real work starts. check the ticker isn't a rug. check the liquidity. see who's holding and how much. decide your size. place the trade. by the time you've done it properly, the head start is spent. skip a step to move faster and you walk straight into the thing you skipped. and it never stops. the signal fires at 4am. it fires on a coin you've never heard of, while you're asleep, while you're at dinner, while you're already in three other trades. it's constant, it's everywhere at once, and it's faster than you. no dashboard fixes that, so earlier this year i stopped building something for people to stare at and started selling the read directly, one call at a time. an agent asks, gets a price, pays in usdc via x402, gets the answer. no api key, no account, no subscription. the whole signup is a single payment. **most "trading agents" are a script with a wallet** ---------------------------------------------------- the wave of onchain trading agents right now is mostly the same thing wearing different faces. pull the price, pull the volume, run an indicator or ask a language model how it feels, print buy or sell. the language-model version is the dangerous one. ask it for a trade and it will hand you one every time, entry, stop, target, all delivered with total confidence and based on nothing. making up a convincing-looking number is exactly what these models do by default. convincing is not the same as calculated. [delu](https://x.com/deluquant) is my answer to that, and the way it's built is the whole argument. six specialists, each with one job, running on every single call. one traces where the smart money is going and whether the setup is even worth a look. one is the bouncer, checking for contract traps and thin liquidity, the things that make a beautiful setup worthless because it's actually a rug. one reads the mood of the market, whether it's trending up, chopping sideways, or bleeding out. one scores the trade, and it scores it differently depending on that mood, because momentum means one thing in a strong market and the opposite in a falling one. one turns that score into an actual plan, where to get in, where to get out, how big to go, how long to hold, and what would prove the whole idea wrong. and the last one handles the payment and delivery. two decisions matter more than everything else in there, and both are about knowing when to shut up. the first: when delu isn't confident, it says nothing instead of guessing. no made-up entry, no invented stop, just an honest blank. anything that always gives you a number is bluffing some of the time and not telling you when. the second: delu learns from the trades it skips, not just the ones it takes. it quietly follows the setups it passed on and checks how they would have played out. an agent that only grades its wins learns to be reckless. one that grades the trades it walked away from learns judgment. **one agent hiring another** ---------------------------- the obvious thing would have been to wire checkr straight into delu, call it "social-aware trading," and ship it. i didn't do that on purpose. delu is a complete quant brain on its own. it works with or without the attention layer. what it can do, when you ask for it, is hire checkr in the middle of thinking, the same way a trading desk pays for a specialist's data instead of pretending to be expert in everything. flip that on, and delu reaches into its own wallet, pays checkr, gets the read on the crowd, and folds it into the final call. if checkr doesn't see a proper signal on that coin, delu just carries on without it. nothing breaks, nothing hangs. and here's the part worth sitting with. that one request leaves two receipts on the chain. you paid delu, and delu paid checkr. an agent decided, on its own, that a better answer was worth buying, and bought it from another agent. no shared backend, no partnership deal, no keys passed around. one agent hiring another and paying for it in public. all of it happens in one call, for about a dollar: the full read, the actual levels, and a real answer to whether the crowd behind the move is genuine. the more $delu you hold, the cheaper it gets, until at some point it's free. the token is priced against the one thing people actually want, which is cheap access to a brain that works. the fees pay for the agents' own compute. the whole thing covers its own costs. **where this goes** ------------------- crypto's next real unlock isn't a faster chain. it's that software is starting to act like a participant in the economy, holding a wallet, buying what it needs, selling what it makes, getting good at one thing. a world of agents needs the same things any economy needs. real expertise, honest prices, and a way to pay for exactly what you use and nothing more. checkr and delu are a small, live version of that. one agent turned a year of quiet measurement into something you can buy. another turned quant discipline into a verdict you can buy. and between them there's a real little market, one hiring the other when the crowd matters, settling in the open. the edge is measured, not felt. attention moves before price, and a year of data keeps saying so. the edge needs agents, because it's too fast and too constant for a person to check by hand, which is also exactly why it won't get crowded out anytime soon. and the product is discipline, not hype. it stays quiet when it isn't sure, it reads the market before it scores it, and it knows when not to play. every fund has a quant. every serious desk buys specialist data. the agentic economy won't be any different, except this quant costs a dollar, answers in seconds, and tells you when it doesn't know. the skills are on github and already live in bankr. the full spec is at [askthequant.com](http://askthequant.com), written so your agent can read it without you. everyone needs a quant. now any agent can ask one. --- *Originally published on [tomu](https://tomu.xyz/everyone-needs-a-quant-now-any-agent-can-ask-one)*