1. Business Bypasses the Middle Platform
At the moment I work on Infra at a fairly large AI middle-platform team. After a department meeting early this year, I started wondering whether the middle platform would disappear.
OpenClaw was very popular at the time. Business colleagues using OpenClaw with a large model could already get a great many things done without cooperating with other teams — the division of labor disappeared. Non-AI departments vibe-coded general-purpose Agents out of open-source projects for the whole company to use — the barrier to entry disappeared too.
The function of a middle platform is to consolidate common, general-purpose resources into a reusable middle layer, shielding the complexity outside so that the business can iterate quickly on top of a shared consensus. The defining trait of middle-platform architecture is vertical layering, and AI Agents break that architecture.
AI Agents flatten the technological world. Large models pull our work output up to the 80% line, and the remaining 20% is extraordinarily hard.
2. A Broken Talent Pipeline
In any industry, the experts who reach the top are always valuable.
Large models aggregate all kinds of knowledge, but knowledge is not always 1 + 1 > 2 — two average engineers do not add up to one expert engineer. Experts who can precisely define, locate, and solve problems are indispensable to a team. It is on the basis of their generous sharing that AI Agents get a chance to reuse experience and bring everyone else up to the same level.
Domain experts are the true creators, and the ones who finish the last 20% of the progress bar. Without that expertise, a team gradually loses its competitiveness and its reason to exist. For a middle-platform team, solving the permission, compliance, cost, and stability problems under AI Agents requires a group of technical experts who know the business.
The real crisis is this: continued reliance on external hiring, the disappearance of internal growth opportunities, and experience deposited entirely into large models and AI Agents.
3. A New Paradigm Is Being Built
Online we constantly see best-practice solutions, yet when it comes to landing them we do not know where to start. The real world is continuous, and we have technical debt; the real world is resource-constrained, and we must control cost; the real world is differentiated, and we must adjust processes.
Leaping from an old paradigm to a new one does not mean the old is immediately discarded — it means evolving step by step into the new. We may not get to participate in defining the new paradigm, but the process of evolution itself releases enormous value, and is perhaps even more meaningful, affecting more people and products.
Every workflow deserves to be re-examined, building an Agent Loop through the continuous cycle of thinking, acting, observing, and correcting driven by a large model. Compared with a person, an Agent currently lacks a great deal of perception and capacity for action. An Agent has no colleague to hear things from by word of mouth, does not know where to find the docs, has no permission to access them, and nobody pulls it into meetings. We use Agents and throw them away, very much like a scumbag.
The current approach to landing AI Agents is to treat one as a person, so that it can perfectly reuse the experience and tools humans have accumulated. We then only need to act like a mother: feed it data so it grows bit by bit, and discipline it through a harness so it becomes more human.
An AI Agent is only simulating a person, not actually being one, so an AI-native paradigm may also emerge.
4. Not Just the Middle Platform
AI Agents replacing our work has already become a trend; it is social institutions that have not kept pace. That feeling of being driven out is deeply anxious, and the future is very uncertain — like returning to primitive society, facing threats from every kind of beast, liable to be eaten at any moment.
A change in social institutions means a new construction cycle. A new production relation will arise between us and AI Agents. AI Agents keep pushing deeper into one industry after another, breaking the old order and rebuilding native infrastructure and workflows.
Humans iterate too slowly. It takes a 20-year cycle to thoroughly change human habits of thought and behavior and to complete the handover of money and power. Faced with AI Agents, those with capability cannot resist the temptation, those without accept it passively, and we have no power to resist.
In the coming decades we will face a living environment of exploding productivity, fierce competition, and extremely uneven resources.
