Vision
Specialised or generalist agents: which will win?
The question is badly posed. What protects a vertical agent is not its knowledge, which the model will eventually have, but the responsibility it accepts.
Grégory Hissiger
Co-founder, Cobalt
Neither, as the question is posed, and that is the one point I would bet on. What protects a vertical agent is not its knowledge, which the model will eventually have, it is the responsibility it accepts to carry, and no advance in models transfers a responsibility.
Best to start by taking the opposing case seriously, because its argument is good and most vertical vendors pretend not to hear it.
The generalist argument, in its strong form
The model improves for everyone at once, and it improves fast. Every piece of knowledge a vertical product has patiently stacked into its instructions turns up, two versions later, in the base capabilities everybody gets for free. Recent software history backs that reading more often than the reverse.
Add distribution, which decides more battles than technology does. Copilot is already installed in every Microsoft tenant, with no purchase to make. A horizontal agent like Viktor raised $75 million in Series A three months after launching and advertises more than three thousand integrations. Against that, a vertical vendor turns up with an ATS connector and a lot of conviction.
And the most uncomfortable argument for us: a buyer prefers one supplier to five. Every specialised product adds a contract, an invoice, a security review and an integration to maintain, and that friction is real even when the product is better.
What that argument misses
It treats business knowledge as a vertical product’s only asset. That is false, and it is the least solid of the three it holds.
The first asset is access. An agent working inside a trade owns write integrations, which is not a matter of intelligence but of error handling, duplicates, permissions and recovery. A model twice as good does not grant you a credential on a client’s ATS, and will know no better what to do when the API errors on the tenth profile of a batch of fifty.
The second is tacit knowledge, and it holds up better than expected because it was never written down. A model learns from what exists as text; what an account manager knows about a client was never published anywhere. That also explains a result we have discussed elsewhere: auto-generated skills produce no measurable gain, while those written and reviewed by humans regularly beat the baseline.
The third is the least discussed and decides the most. A recruitment system will fall under Annex III of the AI Act on 2 December 2027, with the documentation, traceability and human oversight that entails. Somebody must carry that, contractually, for that specific use. A horizontal vendor will not do it for one trade among three thousand, and will be right not to.
The split does not run by industry
Here is the wording I believe is right, and it follows no industry boundary: the line runs where a mistake has a named victim.
Summarising a channel, finding a document, preparing a table, explaining a contract: a mistake there costs a few minutes and is corrected without anyone being harmed. The generalist wins that ground, it has already won it, and a vertical vendor clinging to it is selling a dearer product for an equivalent service.
Writing to a candidate, deciding a profile will not be submitted, updating a record a client will rely on: a mistake there touches an identifiable person who can complain, and sometimes sue. On that ground the question is not which product reasons best but who answers when the answer was wrong. That is exactly why we hold that nothing outbound is automated, and that rule is not an implementation choice, it is the very content of what a vertical product sells.
That line has a merit no sector segmentation has: it cuts across companies instead of sorting them. The same consulting firm belongs to the generalist for preparing its meeting notes and to the vertical for everything touching a candidate.
The test that predicts who survives
It fits in one question to put to any vendor, ourselves included: what would you lose if the underlying model became twice as good tomorrow?
A product answering “we would be much better” has the right answer. What it holds, the access, the process, the responsibility, does not depend on model power, and a stronger model only increases the value of what it holds.
A product that hesitates holds a layer of prompts, and it will be absorbed. There are many of them, they sometimes raise a great deal of money, and they are a good part of why Gartner expects more than 40% of agentic AI projects to be cancelled before the end of 2027. These are not lies, they are bets on a rent that does not exist.
We apply the test to ourselves and accept the consequence. If a model made our business procedures redundant while leaving us the write integrations, the approval line and the hosting, we would be more useful than we are today. If the reverse happened, we would have been wrong.
Where the vertical loses, and it often does
It would be dishonest to conclude that specialisation protects by nature. It protects under precise conditions, and outside them it costs a lot for nothing.
When the process is already standardised and public. A domain whose rules are written in a standard, a framework or accessible documentation holds no tacit knowledge. The model has read it, and the vertical product sells an arrangement of information rather than knowledge. It will be absorbed, rightly.
When the market is too small to fund the integrations. A vertical’s main asset is a set of write connections that are expensive to build and more expensive to maintain. Below a certain market size nobody can pay for that upkeep, and the product decays until a decent generalist is enough.
When the vertical only rephrases. This is the most common and most fragile case. A product interposing an interface and a prompt library between the customer and a model holds nothing: no credential, no responsibility, no process. It lives as long as the buyer does not know they could get the same thing directly, which is a lead that shrinks every quarter.
When the horizontal buys rather than builds. The likeliest outcome for many verticals is neither victory nor disappearance but acquisition, because buying the integrations and the installed base costs less than learning the trade. That is not a failure for the vendor, and it is not neutral for the customer, whose supplier changes priorities overnight.
What the economics say, and it applies to products too
The best indication comes from an unexpected place, employment data rather than software.
The Stanford work on young workers shows the decline concentrating where AI automates work, and staying muted where it augments a person who remains at the centre. We drew out what that changes for hiring juniors, and the same distinction applies to products: what gets absorbed is what was fully describable as a task, what survives is what leaves a person accountable in the loop.
A generalist agent automates. A well-designed vertical agent augments, by bringing the decision to the right person with what they need to settle it. The two will coexist for that reason rather than out of commercial courtesy, and the real question becomes their coexistence: two agents that both know you are not managed like an agent and a tool.
Frequently asked questions
Will generalist agents replace specialised ones?
On the reasoning part they already have, and a vertical product mostly selling good prompts has no future. On access, responsibility and compliance, no: a more powerful model does not give you write permission on a system of record, nor does it sign an impact assessment on your behalf.
What does a vertical agent actually hold?
Three things a model does not manufacture: write integrations with their error handling, tacit knowledge nobody had written down before, and acceptance of regulatory responsibility for one specific use. The first can be copied, the second is paid for in years, the third commits a company.
How do you tell whether a vertical product will survive?
Ask what it would lose if the underlying model became twice as good tomorrow. If it loses most of its value, it was a layer of prompts. If it loses almost nothing, it held access, a process and accountability, and the better model simply makes it more useful.
Should you wait for generalists to cover your trade?
It depends what waiting costs you. On a low-consequence use, waiting is reasonable and free. On a use where a mistake touches a candidate, a client or a contract, waiting means deferring the question of who is accountable, and that question will not resolve itself.