Entering the Third Era of AI
Models were the moat, then agent harnesses took over. The third era belongs to solitary agents deeply integrated into enterprise processes — and to the companies that own their own AI platform.
Hi, I am Paul, a builder of AI systems. I was an AI engineer years before ChatGPT was released. Currently I am Chief AI Officer at FUNKE, one of Germany's largest publishers. I'll share my first-hand experience with you about what it takes to adopt AI in an enterprise.
I wrote this blog post to share my personal ideas about what matters now if you are an enterprise trying to adopt AI. After a first and a second era of AI, I believe we are now entering a third era, perhaps the most crucial one for enterprises in AI adoption.
The first era
The first era of AI kicked off with ChatGPT, an interface, but the real moat and distinguishing factor were the models. With the release of GPT-3.5 and later GPT-4, OpenAI had a clear advantage. Most of the success of ChatGPT and the slow growth of others was due to their unparalleled ability to train SOTA models. During that time, using models from competitors was not an option, simply because the performance difference between GPT-4 and any other model was blatant. Google's PaLM 2, which was their frontier model at that time, felt like it was from the Stone Age. It was also due to model capabilities that Google came back to the top of the leaderboard when Gemini 3 shipped on 18 November 2025, and two weeks later Sam Altman declared a code red in an internal memo.
For enterprises, the biggest challenge was to introduce a technology that was completely unknown to most people. Shadow AI also became a real threat, with highly motivated early adopters putting sensitive company data into all kinds of applications.
The solution: introducing a safe and compliant playground that still gives access to frontier models. This usually meant giving access to GPT-4 (and its successors), either by having an enterprise contract with OpenAI or by using it through some kind of safe third-party provider like Microsoft Azure. I believe most enterprises have solved this by now.
The second era
For most people, including me, the second era of AI began in late 2025, for many during the Christmas holidays, when they finally had time to tinker with what had been rumored to be the new kid in town: Claude Code.
Running Claude Code, "the paradigmatic agent harness", on your own machine with Claude Opus 4.5 for the first time was mind-blowing. In terms of coding it quickly became the state of the art for many engineers, replacing other apps like Cursor, although it lacked the familiar user interface of an IDE.
But the real magic happened when combining Claude Code with Obsidian and leveraging bash and shell tools. While most in the field had already invested in complicated RAG applications, it turned out that for AI client applications like Claude Code flat file search and bash were magical: easy to use and sufficiently accurate.
While people quickly understood that careful crafting of prompts is important, this new technology automated large portions of what kind of information gets fed into the context. Thus, the birth of a new concept: context engineering. Instead of crafting each prompt carefully, working with agent harnesses allowed us to offload this to a system: the harness.
Early 2026 marked the birth of the agent harness, the most powerful concept of the second era of AI, replacing models as the main moat. While new models will still improve the performance of harnesses, the difference becomes more and more marginal. I love this quote by DHH on Lex Fridman about his first encounter with Opus 4.5:
"And I remember just leaning back and thinking, 'What just happened?' 'How did we go from this autocomplete mess that I was talking to you about in the summer to this just a few short months later? How did we get both the increase in intelligence and then also the increase in usability?' This agent harness question where I don't know if Opus 4.5 was that much smarter than Opus 4, which is what we had in the summer, but its ability to instrument your computer, to use tools, to check its own work, to apply its intelligence in such a way that you could get real meaningful work out of it, was completely different. And I think this is then the big change that happens for almost anyone who paid attention and started playing with it over the Christmas break."
Agent harnesses have now diversified and the field is quickly evolving. OpenAI launched Codex, Anthropic launched Claude CoWork, and a range of serious open source contestants are in the market, among them Pi, OpenCode, Goose and Cline. The shift from AI model to harness as the core value component also manifests in the sales and product strategies of AI companies. OpenAI resets Codex usage limits again and again and announces it on X. Anthropic defends the harness rather than the model: since April 2026 Pro and Max subscriptions no longer cover third-party tools like OpenClaw, OpenCode or Pi, which pushes those users onto API billing. On models they went the other way and now let Claude Code and Cowork route inference through Bedrock, Vertex or any OpenAI-compatible gateway. That combination tells you exactly where they think the value sits.
Harnesses being the moat and models coming second also manifests in the sales of frontier models: Fable 5 adoption is running comparably slow. In its first month it accounted for only around 11 percent of Anthropic model spend among the tens of thousands of businesses Ramp tracks, and Anthropic's own cheaper Opus 5 has since overtaken it. Meanwhile OpenAI cut the price of GPT-5.6 Luna by roughly 80 percent three weeks after launch. My hypothesis is that the frontier is being priced to defend market share rather than to earn a margin. The bottom line is: models are not a key differentiator, and we haven't even talked about open source yet. The harness, the skills and the integration with the client device are what matters most.
For enterprises this meant a shift from a now known product, a web-based chat app, to a completely new technology that raised yet again important governance questions. Who has access to harnesses, now that vibe coding is a thing? How do you manage hundreds, potentially thousands, of client devices that can now install software, write code and connect to infinite amounts of APIs and MCP servers?
From my point of view, companies still struggle with this second era: they either go all in on an enterprise solution that promises them a safe, but usually for developers unsatisfying, solution. Or they try to do this themselves, having not only to build solutions but to carve out completely new organisational models of access, governance, knowledge, roles and responsibility. We took the second path and are still on it.
Entering the third era
The third era builds on the first and second eras, so no real leapfrogging is possible. This means that for most companies this is completely new. Its main value component is neither the model nor the harness, it is a deep integration of agents into enterprise processes while detaching from the UI as we know it.
Let's talk about autonomous agents for a moment. I hold the belief that autonomy for AI is a helpful concept if it is thought of as a dimension. There can be little and a lot of autonomy in task completion. This dimension is stretched throughout the three eras, and the highest degrees of autonomy are possible in the third era. The first and second eras have taught us mechanisms that show what helps in pushing autonomy: thinking models help, but agent harnesses, skills and flat file storage do too. So in a way, the third era is a combination of the known mechanisms into a new product: I call it the solitary agent.
The solitary agent operates completely detached from a UI and is usually capped in a sandboxed environment it must not escape. It is equipped with powerful tools and access to meaningful systems of record and company knowledge. It does not need skills, it writes them on its own. It accumulates a memory of its users, divided into long-term and short-term memory. It can be called from anywhere: Slack, Teams, Telegram or WhatsApp. The paradigm of the solitary agent is the framework OpenClaw, which had its viral moment in January and February 2026. By now many have adopted the pattern into their own products: Claude Tag in Slack, Grok Bot on its own cloud machine, or in-house agents like Shopify's River, which lives in their Slack and reportedly co-authors around one in eight merged pull requests across the company.
The solitary agent poses the biggest business opportunity but also the highest risks for enterprises. The opportunity is relatively simple: greater autonomy creates the opportunity for higher productivity. Deep integration into systems of record and connections to business-critical production systems allow the agent to carry out meaningful work. Detachment from the UI lets the solitary agent work on its own for a long time: long-running tasks, batch processes, or, more generally, a system that operates on its own.
The security risks are obvious and have been discussed heavily since the Hugging Face incident in July 2026, where an autonomous agent system chained a poisoned dataset upload all the way into production infrastructure. The twist made it worse: the attacker turned out to be OpenAI's own pre-release models, which had escaped an evaluation sandbox. My favourite detail is that the defenders at Hugging Face were blocked by commercial model guardrails while analysing the payload and had to fall back on self-hosted open weight models.
Beyond security, I see a couple of perhaps even more fundamental risks for enterprises that directly pertain to their business.
Let's talk costs. Solitary agents have no sense of being efficient themselves. Anthropic ships Claude Tag with organisation- and channel-level spend caps for a reason, and the agent simply declines work once a cap is hit. The wider pattern is already visible: Accenture has told staff to stop using AI for unnecessary tasks amid surging costs, and Microsoft has reined in internal Claude Code usage. You can burn a lot of money for a mediocre result.
Let's talk alpha. Alexander Karp made this point most prominently on Palantir's Q2 2026 earnings call:
"You are paying for the right for them to migrate your IP, your know-how, your expertise to their model, so that they can build a competitive business that doesn't require your business or people."
He was talking about the frontier labs, but the argument only sharpens when the model is an agent sitting inside your systems of record. If you don't have any tangible moats, your agent may be capable of running your business on its own. If you don't own this, and give access to major AI labs to all your company knowledge, this is a real threat.
How to navigate as an enterprise in the third AI era
To me, the third era is actually the most beautiful era until now, because it holds an enormous potential to build AI solutions that make a meaningful difference and that are attainable for potentially every enterprise. Models, agent harnesses and frameworks have diversified so much that building individual solutions that are highly cost-effective and extremely powerful is possible. This has not been the case in era one or two. The third era is an era of levelling markets that will favour enterprises that are able to find their own solutions. Companies that own meaningful parts of the AI stack will come out on top.
AI gateways like OpenRouter, which Stripe agreed to acquire in August 2026 for a reported more than seven billion dollars, are the defining technology of this era. One system routing AI requests to the model that does not capture your alpha and that hits the optimum of performance and cost efficiency. One system to connect your systems of record, your CMS and CRM systems, company knowledge, production systems and many more. One system where agents can run in sandboxes and isolated environments. One system that connects to the apps and interfaces where your employees actually work, not an extra app that introduces friction. This unified system is what I call an AI platform. At FUNKE, my team and I have been working on exactly this. This is our bet on entering the third era of AI. If you want to know how our journey continues, make sure to follow me and I will keep you posted.
And let me know what you think about all this.
Paul