The most expensive question on a hedge fund desk is four words long: "why is it down?" A stock the analyst covers drops eight percent before lunch, three portfolio managers ping at once, and the next twenty minutes vanish into browser tabs - the transcript, the news wire, an analyst's note, a podcast someone mentioned, a thread on X. By the time the answer surfaces, the move is half over. Cohesion, a New York company in Y Combinator's Spring 2026 batch, was built by someone who lived that scramble for years and decided a team of AI agents should handle it instead.
Cohesion describes itself plainly: agents that automate research for institutional investors. The product is an agentic teammate for public equity investors - the long/short and long-only funds that pick individual stocks for a living. Instead of a person manually gathering information, Cohesion runs a set of AI agents that monitor the companies a fund covers, track earnings and news as they land, and pull from datasets analysts rarely have time to comb through themselves, including podcasts and posts on X/Twitter. The output is not a wall of raw data. It is differentiated insight: the one line about a supplier comment or a shift in tone on a call that actually changes how a position looks.
Built by people who sat on the desk
Cohesion's credibility starts with who is building it. Devon Krapcho, the co-founder and CEO, spent more than five years as an analyst at Long Path Partners, a fundamental hedge fund, covering software companies. That matters more than it might sound. Tools for investors are usually imagined by engineers guessing at what an analyst's day feels like. Krapcho already knows - he knows which parts of the job are judgment and which parts are just gathering, reading, and summarizing that eat the morning before the thinking can start.
He is not building alone. Co-founder Matthew McBrien came from Amazon, where he worked in security at AWS - the kind of background that matters when your customers are funds handling sensitive positions and want to know exactly where their data goes. The third co-founder, Matt Munns, previously built AI for investors at T. Rowe Price, one of the largest asset managers in the world. Between them the team pairs a decade-plus of public-markets investing experience with engineering that has shipped at real scale. One side knows the workflow cold; the other knows how to build software that institutions will actually trust.
There is a pattern in the best vertical software companies: the founder is a former user, angry about a specific daily annoyance, who refuses to accept that it has to be that way. Cohesion fits the mold. The founding team's public-markets experience is not a line on a pitch deck - it is the reason the product knows the difference between information an analyst wants and noise they will resent. That instinct, hard to fake and harder to reverse-engineer, is what a fund is really buying when it lets Cohesion's agents watch its coverage.
From coverage list to a signal you can act on
The mechanics are easier to picture than most AI pitches. An analyst tells Cohesion which companies they cover. From there the agents run continuously in the background. They ingest the structured stuff every investor already watches - earnings reports, filings, news - and the unstructured stuff most people skip because there is simply no time: an hour-long industry podcast, a founder's offhand comment on a panel, a credible account posting supply-chain chatter on X. The agents read it all, connect it to the positions the fund holds, and surface what changed. No manual setup, no prompt engineering, no dashboard to babysit.
That last step is the whole point. The difference between a data feed and a teammate is that a teammate has read the material and tells you the part that matters. Bloomberg terminals and sell-side research give an analyst more information; Cohesion is trying to give them less of it - the filtered, connected, why-you-should-care version. In a job where everyone has access to roughly the same public facts, the edge comes from noticing the thing the desk across the street scrolled past.
Who is paying for this already
Cohesion is not a demo looking for users. Before any splashy public launch it was live with more than ten long/short and long-only fundamental equity funds, representing a combined ten billion dollars and more in assets under management. For a three-person company still inside its YC batch, having real money managers depend on the product day to day is the strongest signal there is. Funds do not adopt an unproven research tool for fun; they do it when it saves an analyst hours or catches something they would have missed.
The customer profile is specific by design. These are fundamental investors - people who build a thesis on a business and hold it - rather than quant shops running signals through models. That specificity shapes the product. A quant fund wants clean numerical feeds; a fundamental analyst wants context, tone, and the human color of a story. Cohesion leans into the second, which is exactly why podcasts and social posts belong in its diet and not just in a footnote.
Alternative data that nobody had time to open
For a decade the finance industry has chased "alternative data" - credit-card panels, satellite images, web traffic, app downloads - as a source of edge. The dirty secret is that most of it became another subscription and another folder the team never opened. The bottleneck was never access to information; it was attention. A human analyst covering twenty names cannot listen to every podcast, read every filing footnote, and watch every relevant account in real time. Something always gets missed, and in public markets the thing you missed is often the thing that moved the stock.
Cohesion's bet is that agents change the math on attention. An agent does not get tired at 4 p.m., does not cover only its favorite names, and does not skip the boring transcript because the boring transcript is where a quiet guidance change hides. By treating unstructured sources - audio, social chatter, long-form commentary - as first-class inputs rather than afterthoughts, Cohesion tries to make alternative data useful for the first time: not a spreadsheet you subscribe to and ignore, but a teammate that has already read it and flags the sentence you needed.
How Cohesion sits next to the incumbents
Cohesion enters a crowded and fast-moving corner of fintech. A wave of companies is aiming AI at investment research, from document-search tools to analyst copilots. What separates them is usually depth versus breadth: a general assistant that can answer any question shallowly, or a system built around one workflow that it does deeply. Cohesion has planted itself firmly in the second camp, and its founders' resumes are the reason it can. The table below sketches the terrain rather than settling it.
| Approach | What it optimizes for | Where Cohesion differs |
|---|---|---|
| Bloomberg terminal | Universal data access & speed | Cohesion filters and connects, rather than piling on more feeds |
| Sell-side research | Analyst opinions on covered names | Continuous, position-aware monitoring instead of periodic notes |
| General AI chatbots | Answering any question, shallowly | Purpose-built agents for the equity-research loop |
| Alt-data subscriptions | Raw signal you interpret yourself | Reads unstructured sources and surfaces the takeaway |
It also helps that the workflow Cohesion attacks is universal across its target customers. Every fundamental fund, regardless of size or strategy, has analysts who cover a list of names and dread the same overnight blind spots. That means the product does not have to be reinvented for each new customer; the core loop - cover, monitor, connect, surface - travels. For a small team, selling one clearly-defined thing that many firms need the same way is far more scalable than bespoke consulting dressed up as software.
The competitive question Cohesion has to answer over time is durability. Continuous monitoring and clean summarization are valuable, but larger platforms will keep adding AI features, and a general model gets better every few months. Cohesion's moat, if it has one, is the same thing that got it early customers: it is built by people who know precisely what a fundamental analyst needs and precisely what they will ignore. That taste is harder to copy than a feature.
Small team, serious stakes
The business model is the familiar shape of enterprise software sold into finance: a subscription placed in front of funds, most likely priced by seat or by firm, that grows as more analysts and more funds come on. It is a good market to sell into. Investment firms have real budgets, a clear willingness to pay for anything that plausibly produces alpha or saves senior analysts' time, and a low tolerance for tools that waste either. The flip side is that they are demanding customers who expect security, reliability, and discretion - which is where a co-founder out of AWS security earns his keep.
There is something quietly striking about the scale-to-headcount ratio here. Three people sit, through their software, inside investment decisions touching more than ten billion dollars in assets. That is the leverage that agentic software promises and that public markets, with their standardized data and high value per correct insight, are unusually well suited to reward. A small team that nails a painful, high-stakes workflow can matter far out of proportion to its size.
The teammate that never clocks out
Cohesion is early - a 2026-founded, three-person company that has proven demand and now has to prove it can scale the product, the trust, and the team without losing the taste that made it work. The name is a fair description of the ambition: pull the scattered pieces of a research process into one coherent view. If it works, the junior grind of gathering and reading shifts to agents, and human analysts spend their hours on the part machines are worst at - forming a conviction and having the nerve to hold it.
For now, the pitch is refreshingly concrete. It does not promise to pick stocks or replace the analyst. It promises to be the teammate who read everything overnight and is waiting with the summary when the market opens - and, when a position drops eight percent before lunch, to already have the answer to the only question anyone is going to ask.