We track the agentic SaaS landscape from a research perspective — not as competitors, since most of these products are not infrastructure, but as a leading indicator of the questions our customers will be asking us in nine months. The tracker covers sixty-two products that launched between mid-2023 and mid-2024 with some variant of "AI agent" in the headline pitch. As of last month, slightly more than half of them had been renamed, sunset, or absorbed into an acquirer's platform. The capability survived. The product label and, in many cases, the vendor did not.
This is not a prediction. It is the observed half-life of a category. And it is consequential for any enterprise CIO who is evaluating an agentic SaaS purchase right now, because the question on the table is not whether the capability is real. The capability is real. The question is whether the vendor delivering it will exist in the form being purchased eighteen months from now. The base rate, on current evidence, is no.
What follows is what we have learned from watching the cohort, and how to invest in the underlying capability without taking on the vendor risk that has consumed half of it.
Why the half-life is so short.
Three forces are operating simultaneously, and they compound. The first is base-model commoditization. The agentic SaaS that launched in 2023 was differentiated by access to a particular base model. The differentiation has eroded as model providers have converged on similar capabilities at similar prices. A product whose entire moat was "we wrap GPT-4 in a workflow" has had its moat collapse as the wrapping has become a commodity engineering exercise.
The second is enterprise procurement maturation. Buyers in 2026 are asking questions in week one of evaluation that buyers in 2023 did not ask until month nine. Where do my embeddings live, who controls the upgrade path, what happens to my workflows if you ship a breaking change. Vendors that built their product on the assumption that buyers would not ask these questions are losing the deals they expected to close. The procurement maturation is irreversible.
The third is acquisition consolidation. The most successful agentic SaaS products are being acquired by platform vendors that need the workflow surface area but do not need the underlying business. Acquisition is not a death event in the financial sense, but it is a death event for the customer experience: the product is folded into a platform, the pricing is reset, the roadmap changes, and the buyer who selected the standalone vendor finds themselves running on a platform vendor's roadmap.
“We did the procurement work twice. The first time we picked a standalone agent vendor. Eight months later they were acquired and the SKU was discontinued. We re-platformed onto a different agent vendor, who got acquired six months after that.”
— VP Engineering, $310M ARR fintech, install 031
What survives the half-life.
The capability survives. The capability of having a system that retrieves from your corpus, drafts in your voice, executes against your stack, and stops at editor sign-off is not going away. If anything it is becoming a baseline expectation. What is volatile is which vendor delivers it under which label.
Three things, specifically, do not survive the vendor cycle and need to be owned at the architecture layer if the deployment is to outlast the vendor.
The corpus. The accumulated set of artifacts the model has been fine-tuned and indexed against. If the corpus lives in the vendor's environment, it goes with the vendor. If the corpus lives in the buyer's tenancy, it survives any vendor change.
The workflow definitions. The structured representation of the steps a workflow takes, the systems it touches, the editorial gates it stops at, and the policy constraints it operates under. If these are encoded in the vendor's proprietary workflow runtime, they are not portable. If they are encoded in a runtime-agnostic format the buyer owns, they survive a re-platforming.
The institutional judgment. The accumulated record of editor corrections, brand voice tunings, and policy adjustments that turned a generic model into one that produces the buyer's specific output. If this judgment is captured as fine-tune data on tenant-owned weights, it transfers. If it lives as configuration in a vendor's UI, it does not.
How to invest in the capability without betting on the vendor.
The pattern that survives, on current evidence, is to separate the three durable assets — corpus, workflow definitions, institutional judgment — from the runtime that operates on them. The runtime can be a vendor product, an open-source framework, or an in-house build. It is replaceable on a quarterly basis without disrupting the underlying assets. The assets compound. The runtime rotates.
This is the architecture pattern we run at Knyte. The corpus is indexed into the buyer's tenancy. The workflow definitions are written into a runtime-agnostic node graph the buyer owns. The institutional judgment is captured as fine-tune training data against tenant-owned weights. If a different runtime turns out to be better in eighteen months, we swap it. The deployment does not restart.
We are not arguing every enterprise should adopt this pattern from a Knyte install. The pattern itself is the point. Buyers who adopted it via in-house builds, open-source frameworks, or other architecture vendors are also weathering the half-life intact. Buyers who built on a single vendor's full-stack agentic SaaS are, on the evidence of our tracker, doing the procurement work over again.
Three questions to ask a vendor before signing.
- If you were acquired by a platform vendor in nine months, what would happen to my deployment? A vendor that cannot answer this without hesitation has not thought about the question, which means the buyer is the one taking the risk.
- Can my corpus, my workflow definitions, and my fine-tune data leave your environment in a portable format? If the answer is anything other than "yes, here is the export schema," the buyer is locked in regardless of contract language.
- What does month eighteen look like for a customer who deploys today? A vendor that talks only about month one and month three is selling a pilot, not infrastructure.
The half-life is real. It is not a critique of the vendors that did not survive it — many of them built genuinely useful products that were absorbed because the underlying capability was valuable. It is a critique of treating those products as infrastructure when their economics suggested otherwise.
The capability is durable. The vendor delivering it, in this category at this moment, is not. Investing in the former without overinvesting in the latter is the architectural discipline that determines whether your AI portfolio compounds or restarts every eighteen months.