
The death of SaaS has been so heavily covered since the Saaspocalypse that wiped the stock value of several large SaaS platforms drastically, and more individuals are reporting complete SaaS tool replacement coded and driven by their OpenClaw agents. It seems like it’s only a matter of time, right?
That’s the thing, I don’t think so, at least for now. Let’s get into why.
Understanding AI Agents and MCP
Ever since the popular explosion that came from OpenClaw, localized AI Agents have been all the rage. AI Influencers and engineers wiped out stock of Mac Mini’s in several locations so they could build and deploy their new AI Assistants, then find out what they could do. Those that were technical enough were getting some amazing results. Then, the influencers started posting headlines, claiming you could run a whole business with just you and your Mac Mini with OpenClaw as the driving force.
The thing is? Many of them did, or at least their content businesses. They were able to build content gathering, parsing, and suggestion engines using OpenClaw quickly and easily, just using prompts to write the skills. Context was kept based on the markdown files that preserved the agent’s sense of self. Connections to tools came from Model Context Protocol (MCP) servers and connections, first introduced by Anthropic and rapidly became the API replacement for AI Agents. With one connection and a SKILL markdown file, your agent could connect to any number of Agentic tools to extend its capabilities. Now, creating video or audio files were a snap, and could be created through a prompt. My own platform, Wayfinder, opens up creation, review, editing, and approvals to AI Assistants via MCP to allow collaboration in real time while creating courses.
The inevitable question would then become: If AI Agents could connect to any and all tools to automate workflows, why do we need SaaS tools at all? Specifically, why do we need a CRM, ITSM, ITOM, etc. when we can just connect to multiple tools and schedule events, run automations, and through a prompt make things happen?
Then, developers started recreating the tools they had already been working on: CRM, Workflow Automation… One developer even vibe-coded a complete replacement for Palantir over a weekend. Development speed, particularly with AI Agents, meant that companies could just build the software they needed in weeks or months instead of taking quarters to configure and personalize an existing platform.
That’s what really drove the SaaSpocalypse: Not that anything HAD happened, but that anything COULD happen. AI Agents with MCP connections could replace a lot of workflow steps, and whole systems could be generated in days.
Why SaaS Is Not Dead: Complexity of Scale
With such a slam-dunk, why haven’t companies given up their SaaS platforms and built their own with home-coded software? The answer is in the complexity of scale.
Most of the examples above were single developers and solopreneurs building the platform they needed with a full understanding of the context in which they would use it. It works for them, but wouldn’t necessarily work for anyone else. Why? Because they already know what to expect, they have built out their own individual workflow and processes, which their home-grown solutions and AI Assistants understand and follow. When they share it out, including full process and prompts used? Not everyone gets the same results.
One thing that SaaS platforms bring to the table is years and years of domain expertise, built into a platform dedicated to be user-friendly (mostly) and configurable (hopefully), and yet conforms to the industry standard for their domain. Salesforce is the dominant player in Sales because, well, they’ve been around for a while and know how playbooks work. ServiceNow dominates the workflow automation arena because they have, literally, built workflows for thousands of companies with varying processes, and identified the most successful workflows for each of the domains serviced. Workday knows HR. Docebo knows LMS. These platforms are where they are because they know their domain so well, they appeal to the masses.
To compete with this level of domain expertise and dominance, the AI Agent would need to have some serious contextual reference built out, and the time it takes for a single person to build out that context sufficiently to be useful exceeds the dedication of the average user in the space. They would rather have a platform that “just works” than code their own.
Why AI Agents Need Time: Complexity of Tools
Now, given time, an AI Agent could generate enough context to be a reasonable replacement for automation and workflows within a given domain for a person, if they had all the right tools. Sure, it could be possible to build them, but then it would need to use those tools, and that’s a whole new ballgame.
Now, your SaaS platforms integrate together, generally, through APIs. That’s great, that’s helpful, and integrations create more usability, though also more complexity. For AI Agents, that would mean MCP connections.
Now, MCP connections are growing, but not yet ubiquitous. In fact, most of your larger platforms likely do not have MCP connections, because their development cycles are often quarters or years long. That, and the majority of their users do not yet have AI Assistants using MCP that would connect to their platforms. Why? Because businesses are still trying to figure out how AI is going to work in the workplace, and no one (yet) trusts OpenClaw because it had some very glaring security issues early on.
First Bottleneck: Agent Deployment
So, what’s the bottleneck? Right now, it’s the proliferation of AI Assistants in the workplace. There are not many, I would warrant. Security teams need to be involved because of the level of access these agents have on computers, and therefore network systems. Then, there’s the use case. Right now, AI Assistant configuration and training is very individual and takes time. Most companies are looking to turn an ROI on AI in the quarter, and there’s not much tolerance for losses right now. At least, not anymore. Boards and the Street want to know that AI is making money.
How will that change? Assistants will be built into operating systems. In fact, I fully expect Apple’s WWDC announcements on Siri to be just that: Opening Siri to MCP connections, system files, etc. to add much of the OpenClaw functionality into the Mac directly without the security headaches. They would even allow for multiple models to be connected, using Gemini as the default, so you can get just as complex in how your build your Agentic team as you wish. Microsoft will do the same with Copilot in Windows. Linux might just keep with Hermes/OpenClaw/etc. based on the popular open-source solutions at the time.
Now, you’re not installing an Agent, it’s already there. For Apple, that’s 3+ billion devices that could be AI Agent-ready in various forms, all tied to a single login. Windows, I’m sure, will be 5-600x that number. At this point, companies don’t need to decide if they are going to have Agents on employee computers, but rather how those Agents will be used. Agent bottle-neck will be solved.
Second Bottleneck: Tools and MCP Configuration
Okay, so now you have every single company computer enabled with a built-in AI Agent, ready to do work like a boss. What tools do they use? How will they use those tools? How will they discover how to use those tools? If they are external, how do they set up MCP connections?
One of the biggest problems with launching any Agentic AI initiative is access to the tools your employees need. They don’t often know what tools they need or how often (sometimes only once a year), and wouldn’t know how to configure it themselves if they needed. I get it, I really do. If you are heart surgeon, I’d rather you spend all your time learning the latest and greatest on heart muscular function than trying to vibe-code your way through setting up an agent to filter out your email.
Not everyone will be, or should need to be, technical enough to set up specific tools for their system. This could easily be addressed using context records that are generic for a role (much like access roles and instructions for setup), which can then be added to the Agent at startup for a user, making the Agent ready day 1. Wayfinder, by the way, attempts to do this with an AI Knowledge file per role. ^_^
Why SaaS Still Rules
So, now we know that setting up Agents is no walk in the park. That’s why SaaS platforms still have a place: They already have context that’s been researched and identified during the deployment phase. They contain the workflows and processes that were already established prior to the emergence of AI Agents, to be fair, but still solid enough that an AI Agent built into the platform can follow it and make appropriate decisions before checking in for approval. SaaS platforms have the deep domain experience to remain on top, even as the scrappy AI Agents start clawing their way to the top. Which is why they won’t be going away soon.
But, what about the future? I hear you ask.. Good question! Slowly, you will see Agents taking over as context becomes easier to generate and mass-produce. Initially SaaS companies will provide Agent connections, so agents can trigger the SaaS workflows and access SaaS results, rather than users connecting. This will speed things up significantly enough that keeping the SaaS will remain a priority in the short to mid term.
Will it remain that way? Given the speed at which Frontier models are moving, I’d say no, not in the long term. Eventually context, workflows, processes and procedures will all be housed within Context portfolios that are designed to be consumed by the AI Agent first, then the human using it. Agents will have enough information that, even when starting a new job, they can replace the previous knowledge files and update with the latest role and be ready before the employee has finished their onboarding. At that point, with MCP connections being created and generated by onboarding teams, connections built during orientation, and access to the shared corporate context with role-specific context populated per user will be the expected onboarding experience, no one will need to access a SaaS portal. The SaaS will be the data warehouse, with access through MCP and access controls dictating content availability.
I’m not arrogant enough to even attempt to predict when this will happen. There are a lot of details that need to be worked out. What I do know? AI Agents will drastically change how we work, and there will be a growing, need for a parallel Agentic AI-First internet to drive that shift. How existing human-first tools adapt will determine if they dwindle on the vine, or become the AI Agent’s go-to warehouse for domain-specific information. Until then, keep your browser shortcuts. ^_^



















