Cohesion

Cohesion sells the agents' work and nothing underneath it: continuous monitoring of a coverage list across sources that have no structured feed, podcasts and social posts among them, with a judgment about which developments matter to a stated thesis. Remove the models and nothing remains but a list of tickers. The company's own framing is that analysts spend hours a week doing this by hand and the agents now do it for them.

Autonomy is the selling point and is stated as such, agents that monitor and generate differentiated insights without the analyst driving each step. The consequence is contained because the output is research rather than an executed decision, and a portfolio manager still sizes the position. What is missing is any description of what the analyst sees to check the work, whether an insight carries its source, or how a claim the agent got wrong is caught before it enters a thesis.

A monitoring product has two failure modes and neither is measured publicly. The miss is a material development the agents did not surface, which is invisible to the client by definition, and the false signal is an insight drawn from a mis read transcript or a low quality post. No precision or recall figure, no grounding or citation guarantee, no evaluation method and no human review step is published. For a product whose output feeds position sizing, the missed event is the more consequential of the two and the harder to detect.

The claim is more than ten long and short and long only funds live, with a combined $10 billion or more under management, self reported by a company founded this year with three employees. Assets under management behind a client list describe the caliber of the buyer, not the performance of the product. No fund is named, no accuracy or time saved figure is published, and no independent evaluation exists. That is a reasonable position for a company at this stage rather than a criticism of it.

The question that decides this product for an institutional buyer is whether one fund's coverage, questions or thesis notes inform what the system surfaces to another fund, since the customers compete directly for the same alpha. Nothing published addresses the boundary, offers an opt out or states whether client interactions train anything. Source licensing is equally unaddressed, which matters when the inputs include podcast audio and social platform content.

Consumer data exposure is absent, since the material is public company information. The sensitive asset is the client's own coverage list, because what a fund is watching and which theses it is tracking is itself signal. Nothing published describes how that is segregated, retained or protected, and there is no public privacy or confidentiality statement.

There is no public certification, attestation or trust page. One founder's background is in a major cloud provider's security organization, which is a reason to expect competence and is not the same as a control a buyer can verify. At three employees no attestation would be expected yet, and institutional vendor review will ask for one before long.

No licence is required to sell research tooling and none is claimed, which is the correct posture. Two regulated edges go unaddressed. Systems monitoring non traditional sources for signal on covered companies sit near the handling of material non public information, and generated research delivered into an investment process may fall inside a firm's supervised records and research policies. Neither appears in any public material.

The relevant exposure is source selection rather than protected class fairness. An agent that ranks podcasts, social posts and press by apparent relevance inherits whichever voices are loudest and most indexed, which systematically favours widely covered large capitalisation names and English language sources, and thinly covered or non English issuers are exactly where an information edge is supposed to live. Nothing published describes source weighting, coverage limits by language or market, or how the system handles a company nobody is talking about.

The loss from a missed disclosure lands entirely on the fund. No warranty, service level, correction process or allocation of responsibility appears anywhere, and nothing published describes what happens when the agents miss something that moves a position or surface an insight that turns out to be wrong. Research vendors usually leave that risk with the client, and it matters because the output is used in decisions with money attached.

Nothing names a model, a provider, an infrastructure dependency or a data source agreement. A three person company shipping agentic monitoring in 2026 is almost certainly running on third party foundation models, so the practical question for a buyer is which ones, what happens on deprecation, and whether client coverage data crosses into a provider's environment. None of that is answerable from public material.

Nothing published names an integration. No order or portfolio management system appears, and neither does a research management platform, a market data provider or a note taking or collaboration tool. The trial is described as a ticker list sent to the vendor rather than a connection made to a client system. For a product meant to sit inside an analyst's daily workflow, that is the gap that matters as the product matures.

Delivery is a hosted platform and nothing further is stated. No region, tenancy model, private deployment option or residency commitment is published, which is a live question for any fund with data handling obligations attached to its own coverage information.

No pricing is published. The trial offer is concrete: a prospective fund sends twenty tickers and gets coverage through an earnings season. That tells a buyer the unit of consumption is the covered name, even though no number is put against it. It is still well short of a rate.

Deliberately narrow, and the narrowness is the strategy. The buyer is the fundamental equity fund, long and short or long only, and the product is built around how a covering analyst works rather than around an institution's data estate. Nothing indicates coverage of credit, macro, private markets, or the sell side, and no geography beyond the domestic listed market is described.