[July 2026 edition]
The Skill of Not Building in the AI Agent Era (1) — Third Parties Swallowed by the Platform
Published: Jul 25, 2026
Reading time: ~6 min
Building an AI agent, the areas you have to consider are countless. How to give it long-term memory, and how to implement that. How to coordinate multiple agents. How to optimize prompts per model. How to repair broken JSON. How to split the context, summarize it, and re-inject only the relevant parts. Whether to build a model router and a semantic cache to hold down API costs.
All of it is technically interesting. But is that implementation really necessary? Put more sharply: by the time it’s finished, will anyone still need it?
Build the best AI agent you can, and the biggest bill comes due right after. You chase each new model generation, patching what you built until the upkeep outruns you or the model is retired — and the moment you pull out, all that code is just garbage. Those hours cost far more than any API bill. This series tests that view over five parts. Part 1 looks back at what has already happened: how the services that sprang up in the early generative-AI boom were swallowed into flagship features.
Companies Survive, Their Edge Fades
Early in the boom, a flood of services appeared to cover what general-purpose models were bad at. Services to read PDFs, to turn text into slides, to sell prompt packs, plugins to connect external APIs to ChatGPT, services to record conversations as personal memory.
Not all of them vanished. But much of their core functionality is now built into ChatGPT, Claude, and Gemini as standard. Surviving as a company and keeping the feature that once made you viable as an independent technical asset are two different things. Many of these companies, having lost that edge, were forced to pivot — into a different specialty, or a different kind of output.
How Long Can PDF Chat Alone Compete?
ChatPDF, AskYourPDF, PDF.ai — services where you upload a PDF and ask questions — were emblematic products of the early boom. The services still exist. But general-purpose AIs like ChatGPT and Gemini now handle document upload, document understanding including images, and long context as standard. Google’s Gemini API advertises understanding documents, images, and video directly from inputs on the scale of millions of tokens.
“You can upload a PDF and ask questions” alone is no longer a technical advantage for a standalone service. To survive, you must solve problems above mere PDF chat — contract comparison, evidence management, statute cross-referencing, rigorous citation, access control. Not every PDF service failed. But the asset value of the technology that once made PDF chat possible has been lost for good.
Weeks to Build, an Instant to Collapse
Tome, which turned text into slides, was one of the symbols of the early boom. But the company cut roughly a third of its staff in October 2024 as it moved away from presentations, announced the sunset of its slides product in March 2025, and shut down the presentation product on April 30, 2025. Users’ slides that hadn’t been exported were deleted. The team now builds Lightfield, an AI-native CRM.
The better a model’s text, image, and layout abilities get, the easier presentation generation is to absorb into existing creative environments — Microsoft, Google, Canva. A dedicated service discovers the market; a large platform catches up by making it a standard feature. It has always happened in software markets. But in AI, the stretch from being locked onto to being shot down is brutally short.
Implementations Left Stranded Within a Year
When ChatGPT Plugins were announced, many developers rushed to connect external services to ChatGPT. It took more than exposing one API: an OpenAPI definition, authentication, plugin-specific metadata, a publishing method, usage descriptions, and testing.
Then, in February 2024, OpenAI announced it was winding down the plugins beta; after March 19, 2024 you could no longer use new plugins or use one in a new conversation, and on April 9, 2024 existing plugin conversations stopped working. The center of extensibility moved to GPTs and Actions, and the developer-facing successor path became the Responses API in March 2025.
A platform has always controlled its own spec — that has never changed. What’s new is the speed of the turnover. Plugins fell from the spotlight within about a year of launch, and all the platform-specific knowledge, implementation, and verification became legacy at once. When a spec’s life is this short, it vanishes before you can recoup the time you sank into it.
The Supplier Is the Biggest Threat
What’s easy to miss is that the platform is not just a model supplier. ChatGPT and Claude aren’t chosen for their model performance alone. Conversation history and memory, file and tool integration, a refined feel — the polish of the application keeps improving day by day.
Its reach doesn’t stop at the chat window, either. Web and mobile apps, desktop apps, CLI agents that run in the terminal, agents that run autonomously in the cloud. The platform is moving to own AI-that-runs-everywhere, wholesale. Wherever a third party chooses to operate, it runs into the platform somewhere across that spread.
So a third party thinks it is competing on the model, when it is really competing against the platform’s app itself. And the platform builds both the model you depend on and the app you compete with. It can see, from how its whole user base uses the product, what features are in demand, and it has far more engineers and capital to put into that app than any third party.
An AI-app implementation that leans too heavily on one particular model — the moment the platform folds the same feature into its standard product — has no other provider to switch to and, still dependent, is pushed out of the market.
The Cost of a Wrong Investment
PDF chat, AI slides, plugins. In each, a dedicated service filled a gap first, and the platform then swallowed it as a standard feature. And each time, the implementation built to support it, and the time that went into it, see their value vanish along with the business. Simply keeping pace with how fast AI moves takes real effort, and most of the time you’re thrown off before you catch up. The opportunity cost that piles up in the meantime, given enough time, dwarfs whatever you lost on the investment itself.
And AI obsolescence runs one layer deeper than ordinary software. When a library ages, you migrate to a newer one and move on. But in AI, if the API you built on breaks or is retired, your code has to be rebuilt or ripped out. And the smarter the model gets, the more the platform pulls the third party’s entire business domain into itself, like a black hole.
Part 2 examines this two-layer structure — where the implementation tech goes obsolete before the service — through three mechanisms: API retirement, the elimination of surrounding code by improved model ability, and the shrinking of heavy long-context machinery by expanded context windows.