--- title: "Why Agent as a Service Needs Preconfigured Specialists" description: "Traditional SaaS is not vanishing overnight, but DIY AI tools and a flooded market are turning niche software into a red ocean. The next product to design is a specialized, preconfigured agent as a service — plug-and-play for a job, not a generic employee you have to train..." publishDate: Tue Oct 06 2026 00:00:00 GMT+0000 (Coordinated Universal Time) author: Romain Penchenat language: en tags: ["agent as a service","saas product strategy","agentic product design","platform as a service","product design"] canonical: https://blog.romain.garden/blog/agent-as-a-service-preconfigured-specialists/ estimated-read-time-in-minutes: 11 --- Traditional SaaS is not vanishing overnight, but DIY AI tools and a flooded market are turning niche software into a red ocean. The next product to design is a specialized, preconfigured agent as a service — plug-and-play for a job, not a generic employee you have to train yourself. Scroll LinkedIn and you will be told SaaS is dead. That is not true as a cliff edge. Software as a service still means what it has meant for years: a web platform or an app sold on a recurring subscription, with the user’s data hosted in the cloud, and interfaces that let someone solve a problem. What *is* true is a trend that is getting harder to ignore. The interesting **SaaS product strategy** question is no longer whether another niche tool can exist. It is whether you should still be designing one — or whether the job has moved to **agent as a service**. *** ## Why SaaS Product Strategy Now Faces a Red Ocean > **Building another specialized SaaS is still possible, but the market is a red ocean: AI-built substitutes on one side, an oversupply of lookalike products on the other.** The explanation you hear everywhere is the first force. Anyone can now recreate a specialized tool relatively easily with AI. I have watched people ship a CRM dedicated to podcast production with AI and Notion. I have watched people stand up a web app to manage a video-editing process. I have watched people build a multi-source feedback tool, again as a custom web app, fitted to their own needs. Products that used to be something you bought from someone who had already built them are, in those small niches, less attractive than they were. A video-editing process manager is already a thin market. Inside that thin market, a slice of people will now make their own app instead of paying for yours. That explanation is not the whole story. A personal solution is perfectly tailored to your process, and that is a real advantage. It also comes with the bill: you maintain it, you evolve it, you fix the bugs yourself. AI-built personal apps will not replace everything, and the larger the company, the harder it is to self-build the stack. That still does not save traditional SaaS. It simply cuts into the part of the market that used to have no alternative. The second force is, for me, more serious. The SaaS shelf is being saturated by vibe developers. Look at mobile: thousands of new apps have shipped in recent months, and Apple has started refusing entire categories by default — to-do lists, habit trackers, the classic productivity utilities — unless you can prove something differentiating. There are simply too many. That was already true for utilities. It is becoming just as true for the SaaS products we already know. On the web, the same flood shows up as probably millions of new micro products: open source, free tools, tiny services. Creation got so cheap that plenty of people do not even bother to monetize. The hard problem is no longer making the product. Almost any product is findable as open source somewhere. The hard problem is emerging from a mass that keeps getting larger. That is a red ocean. You can still enter traditional SaaS. It used to be relatively easy. Today it takes a lot of communication work and a lot of differentiation. *** ## Platform as a Service Is the Stack Agents Build On > **The products with a durable seat are platforms that hide infrastructure — databases, hosting, payments, mail, analytics — so humans and agents can build on them instead of reinventing them.** While SaaS slowly loses oxygen, **platform as a service** is the first thing filling the gap. It is the infrastructure both of those moves run on: when you build a personal product with AI, and when you want to ship something for other people, fast. The technical layer is already familiar to vibe developers. Supabase makes databases easy. Vercel is the equivalent for hosting sites and products. Stripe sits in the same bucket for payments. Metabase is the one I would flag for structuring and combining data sources so an agent, or a custom tool, can pull from them. Infrastructure used to be specialist work, something only developers touched. It is drifting into the common toolkit of anyone who builds their own tools. Building new platforms that help people ship faster, or that give agents data sources they can actually use, is a real product space. Then there is a quieter category I would call quasi-platforms. Notion started as notes and lightweight databases for humans. The side step they have taken is to keep talking to people through a classic interface while also becoming a platform for agents: a place to store, structure, and hand knowledge between agentic work and human work. Analytics products such as Amplitude used to sell the interface you opened to read the numbers. Part of what they sell now is storage and cleanup, so an agent or a custom tool can tap that data. Even if nobody opens the Notion or Amplitude UI tomorrow, agents will still need those products to organize knowledge and to hold analytics. Gmail is the same pattern in mail. You may never need the Gmail interface again. The product is so present, and so easy to connect through an API or an agent, that you can keep using it from a mail client you built yourself — or through an agent that reads, sends, forwards, and files. The common thread across all of these is that they were technical products first, simplified for humans second, and are now being optimized for use by an AI agent. Once an agent can reach mail protocols, databases, payments, and analytics without reinventing them, it can spend its budget on the value it produces for you. Rebuilding email storage, or standing up and structuring a database, every time you wanted a custom app or a custom agent, would be hell. The platform absorbs that complexity. That is why PaaS has a long runway, and also why there is not that much room left to take. The point of these tools is to become the default, the hidden infrastructure. Once a generalist database works, it is hard to stand next to it. The tools are relatively interchangeable, and you do not need fifteen databases or fifteen mail stacks. There is still a question worth digging into: can we design better platforms, specifically for this era of personal product creation and agentic work? Maybe. It is worth the exploration. It is not the question I find most interesting. *** ## Agent as a Service Turns a Chatbot Into an Employee > **Agent as a service is not a chat that answers in minutes. It is a generic virtual employee with tools, a work loop, and something close to its own computer.** The next step is **agent as a service**. The acronym is uglier than SaaS or PaaS. The products are starting to exist. The first wave was ChatGPT, Claude, Gemini, and their peers: large language models that answered in text, with limited reach into the outside world and limited working time. They would think for a few seconds or a few minutes and hand you a reply. They were not autonomous on long tasks. What has shown up in recent months is closer to a hired generalist. Grok Bot from SpaceXAI, Meta’s Muse — which made a lot of noise in the United States and is not available in France — and ChatGPT Dots are the ones on the table. They behave like generic virtual employees, ready to work with your tools, connecting to them directly, relatively autonomously. They are still language models, but they get a set of tools and a work loop on something close to a personal computer. Inside that loop they can generate and execute the kinds of tasks a human could also do, including work that is extremely long and extremely complex. They are still generic. You are the one who turns them into the employee you actually want: a precise objective, connections to the right tools, maybe routines so that every day, every week, or whenever a given event happens, they react in a given way. That configuration is the work. Done well, it helps your day, helps your job, or replaces some of the ancillary functions around your job. *From subscription software to agents you still have to train — click each step* 1. **Software as a service** — A web platform or app on a recurring subscription, user data in the cloud, and interfaces that let someone solve a problem. Entering that market used to be relatively straightforward. 2. **DIY tools and a flooded shelf** — AI makes it easy to recreate niche products, and vibe developers have filled mobile and web with lookalikes. Apple now refuses undifferentiated to-do and habit apps by default. Creation is cheap. Emerging from the mass is the actual problem. 3. **Platform as a service** — Supabase, Vercel, Stripe, and Metabase hide databases, hosting, payments, and data. Notion, Amplitude, and Gmail keep a human interface while becoming the store of knowledge, analytics, and mail that agents connect to. 4. **Generic agent as a service** — Grok Bot, Muse, and ChatGPT Dots are generic virtual employees: tools, a work loop, and something close to their own computer. You still supply the objective, the connectors, and the routines that turn them into the employee you want. *** ## Specialized Agent as a Service Is the Design Opportunity > **If consumer AI won by needing no setup, the product opening is a preconfigured specialist — not another generic agent you have to train.** What made today’s AI products succeed is that they were simple to pick up. You wanted a recipe, you asked for a recipe. You wanted to know how to do a piece of work, or you wanted a piece of information, and it worked. There was no configuration. Generic agents demand that configuration. You have to think through how you want the thing to work so it can actually fill its mission. That is the opening: is there room for specialized, preconfigured **agent as a service**? Take video production. Will a generic agent be enough, or do you have to teach it how production actually runs, give it tools that work with particular formats, and wire triggers into the workflow so that when rushes arrive it structures them, moves them from one piece of software to another, and runs QA and review on its own? A lot of what used to live inside a production SaaS — the thing that let you do the job immediately — now has to be taught to an agent and shipped already configured if you want it to be plug-and-play. Today you would still hand that agent skills, connectors, and a whole event map for the moments you want it to react. Tomorrow, the product might be a turnkey agent for a director. The same question applies to recruiting, to accounting, to shopping. People may well go to the agent that asks for the least work, instead of rebuilding and reconfiguring a generic one end to end. That is **agentic product design**, and it is not mostly an interface problem. The interface slice stays relatively modest. The real design work is: - Which tools you give the agent, and which context - Which routines it should run, and which events it should react to — how, and when - Whether regular tasks get pushed onto that loop - What format the output and the review take - How it communicates with the user, and which artifacts exist for collaboration, feedback, and review between the two of you You are designing two relationships at once: how the human collaborates with the agent, and how the agent interacts with the world. *Generic employee vs preconfigured specialist* | | Generic agent as a service | Specialized agent as a service | | ------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Setup | You train it into the employee you want — A precise objective, connections to your tools, and routines for daily, weekly, or event-driven reactions. | Plug-and-play for a job — The domain knowledge that used to live in a SaaS is already taught and shipped with the agent. | | Tools and triggers | You supply skills, connectors, and event wiring — Nothing reacts until you decide which events matter and how the agent should answer them. | Tools and workflow triggers ship with the agent — A video agent, for example, would already know formats, structure incoming rushes, move files between software, and run QA. | | What you design | The user does the configuration — The product is a capable generalist. Fit to a role is homework. | Tools, context, routines, and review artifacts — Interface stays relatively modest. The product is how the human collaborates with the agent, and how the agent acts in the world. | **Quick check** If consumer AI won by needing no configuration, what is the stronger product bet for agent as a service? - Keep shipping generic agents and let each user reconfigure them end to end - Ship specialized, preconfigured agents that are already wired for a job **(correct)** - Replace platforms such as Gmail and Notion with a new interface for every task - Double down on niche SaaS, because creating the product is still the hard part Generic agents still ask you to supply the objective, the tools, and the routines. The opening is a specialist that already has the domain, the connectors, and the review loop — the same promise the old SaaS used to make, moved onto an agent. *** ## Designers Stay Useful by Using Agent as a Service > **The category is still being invented, so the useful move is to use the generic agents that exist, customize them, and start putting specialists in front of users.** This is still abstract, because the category is being invented in front of us. Grok Bot, Muse, ChatGPT Dots, and the rest have only been somewhat concrete for about two months. These first generic consumer agents are just arriving. That is exactly when it is useful to start exploring — and, maybe, to start designing the specialists. I am not treating this as a locked prediction. It is a way to look at a possible next market. The practical move is the same one I am making: put your head in it. Use agents. Customize them to automate a slice of life or a slice of work, not only because you need it, but as watchfulness — to understand what they are and how they work. Muse is not available in Europe. In Europe you can test Grok Bot or ChatGPT Dots. If you want the technical picture of how these agents actually loop, I recently followed a Vercel course on that model; it is a useful way to get your hands in the mechanism. For now, that is most of what a designer can do: look closely, test what is on the market, and start putting things in front of users to see what makes sense and what does not. That is how you stay useful in a world that is rearranging itself under your feet. *** ## Final Thoughts Niche SaaS is not forbidden. It is just expensive in a way it did not used to be: you are shouting over a flood, while a share of your buyers can now build a private substitute. Platforms will keep sitting underneath, as the default pipes for data, mail, payments, and hosting. The product that is still being designed in the open is the agent you do not have to train from zero. If there is a craft shift here, it is that one. Less time on another dashboard for a crowded job. More time on the tools, the context, the routines, and the review formats that turn an agent into someone who can actually do the work with you. *** ## FAQs ### Why is traditional SaaS becoming harder to enter? Two forces: people can now recreate specialized tools with AI instead of buying them, and vibe developers have flooded mobile and web with lookalike products, so the hard problem is emerging from the mass rather than creating the product. ### What is platform as a service in an agentic product landscape? It is hidden infrastructure — databases, hosting, payments, mail, analytics — that humans and agents build on so they do not have to reinvent email protocols or database setup. Examples include Supabase, Vercel, Stripe, Metabase, and quasi-platforms such as Notion, Amplitude, and Gmail. ### How is agent as a service different from ChatGPT, Claude, or Gemini? Those first-wave models answered in text over seconds or minutes, with limited reach into the outside world. Agent as a service products such as Grok Bot, Muse, and ChatGPT Dots act more like generic virtual employees: they connect to your tools, run a work loop on something close to their own computer, and can take on long, complex tasks. ### What would a specialized, preconfigured agent as a service include? Domain knowledge, tools, and workflow triggers shipped with the agent so it is plug-and-play for a job — for example a video agent that structures rushes, moves files between software, and runs QA — plus defined output, review, and collaboration formats with the human.