The harness half-life: AI is now moving faster than you can build for it
The machine-learning era took eight years. The agentic era has taken about twenty-four months. What that compression means for builders, boards and investors.
I have been building with AI since my Creditscript days, and I now build agentic systems for clients. The thing I find hardest to explain to boards is not what the technology can do. It is how fast it is moving. The pace is no longer linear. It is parabolic.
Eight years, then twenty-four months
The way I date it, the age of machine learning ran from 2016 to 2024. It started with AlphaGo beating Lee Sedol in March 2016. In 2017 came the paper that introduced the transformer. GPT-1 and BERT followed in 2018, GPT-3 in 2020, and ChatGPT reached the public in November 2022. That was eight years from a research milestone to a technology every company had to have a view on. It was fast by any historical standard, but it was still slow enough to plan around.
The next era began in late 2024. OpenAI released its first reasoning model in September 2024. Anthropic published the Model Context Protocol in November 2024, giving agents a standard way to use tools and data. DeepSeek-R1 showed in January 2025 that frontier reasoning could be reproduced at a fraction of the cost. By spring 2025 there were coding agents that work in a real codebase, and Google’s agent-to-agent protocol for agents to talk to each other. Within roughly twenty-four months, AI went from answering questions to doing multi-step work.
Eight years, then two. If the curve holds, the next shift of the same size will not take two years.
By the time you build the harness, it is out of date
An agent is a model plus a harness. The harness is everything around the model that turns a clever model into a reliable worker. I have built several, and they all have two layers. The first is scaffolding: prompts, tool wrappers, retries and workarounds for the model’s weaknesses. The second is the context and memory engine: what the system knows about the business, and how it assembles the right knowledge for each task. In every harness I have built, the scaffolding was partly obsolete within months.
A newer model would arrive and do natively what my scaffolding had been working hard to force it to do. A protocol would standardise something I had built by hand. A weakness I had engineered around would quietly disappear. The scaffolding did not fail. It was overtaken. The context and memory did not date at all. Each new model simply made better use of them.
That is the real cost of parabolic speed. It is not that you build the wrong thing. It is that you build the right thing for a world that has already moved on.
What lasts
So the value sits in the harness, but not in all of it. It sits in the context and memory, which are becoming the core intellectual property of a company. Here is what has held its value in my own work.
- Your context and memory. What the system has learned about how your business works, and how it puts that knowledge in front of the model, outlasts every model. Models are rented. Memory is owned.
- Your data. Proprietary data and the rights to use it feed that memory.
- Your evaluations. A good test set that measures whether the system does your job well lets you swap in a new model in a day rather than a quarter.
- Your controls. Permissions, audit trails and security are what let a regulated business use agents at all, and they carry over from one model to the next.
- Disposable scaffolding. Build the scaffolding to be thrown away. Keep it small, keep the model replaceable, and assume you will rewrite it within a year.
What this means for investors and boards
In diligence I now ask one question before any other on an AI company: how long would it take you to replace the model underneath the product? If the answer is months, the company carries a depreciation schedule that is not on its balance sheet. If the answer is days, the company owns something durable: its context and memory, its evaluations, its customers.
Boards should apply the same test to their own AI programmes. Fund the scaffolding as if it were software with a twelve-month life. Put the long-term capital into the parts of the harness that compound: context, memory, evaluation and controls.
The machine-learning era rewarded those who built early. The agentic era rewards those who can rebuild quickly.