Recently, I had a drink with some friends working on AI applications, and as we talked, we all sighed.

It's not that the technology isn't good enough; it's mental exhaustion. Last year, we were still obsessing over how to write prompts like incantations. Early this year, we were busy putting harnesses on agents, afraid they'd delete the database. By June, the whole world was caught up in Loop Engineering, eager to turn every workflow into a perpetual motion machine.

But when you actually get a perfect Loop running, watching it tirelessly self-correct, submit code, and pass tests 24/7, a thought creeps into your mind:

What's next?

If we view AI engineering as a geometric ascension game, we seem to have cleared the 'volume' level. When the closed loop of single-agent intelligence is pushed to its limit, what is the next thing that can excite the industry for another half-year?

My answer might be a bit abstract: the Field (域).

To understand why it's the 'Field,' we need to look back at how we moved step by step from 'point' to 'volume.' This isn't just technological iteration; it's a shift in our role in the face of AI.

Point: Prompt is a 'single-point breakthrough' and a 'misalignment of expression'

In 2023, we were 'conversationalists.'

The question was simple: how to make the model understand human language? Chain of thought, few-shot, role-playing... all energy was focused on that moment of input. The Prompt was like a 'point,' and we tried to leverage the entire model's capability with a single sentence.

But this thing had a fatal flaw: the real world is not an isolated question but a tangled mess of information. No matter how sharp that point was polished, it couldn't hold up once the task got complex.

Line: Context is a 'linear extension' and a 'window game'

In 2025, Shopify CEO Tobi's statement, 'Context Engineering is the art of curating what goes into the limited context window,' woke everyone up.

The bottleneck was no longer 'how to ask' but 'what to show the model.' RAG, memory systems, and historical documents became standard. Context stretched the isolated 'point' into a 'line,' giving the model a sense of cause and effect.

But we became like information secretaries, carefully trimming, sorting, and compressing, afraid that giving too little would make the model guess wildly and giving too much would make it lost. This line solved 'accuracy,' but it was still static. The model gave an answer, but who would verify it? Who would execute it? Who would be the safety net?

Plane: Harness is a 'planar constraint' and 'deterministic anxiety'

In early 2026, agents started touching production environments, and disaster struck. Infinite loops, permission overreach, tool call failures... we finally understood that having a smart brain isn't enough; you also need a safe cage.

Harness is this 'plane.' Permission control, log tracing, guardrails, automated testing... it no longer cares what the model says, only what environment the model operates in.

We shifted from conversationalists to foremen, building scaffolding and laying safety nets to ensure the AI intern wouldn't bring down the company. But Harness was still one-off. A task came, it was done, and it ended. Next time, a human still had to press the Enter key. The human was still the Loop itself.

Volume: Loop is a 'three-dimensional closed loop' and a 'serial ceiling'

So by June of this year, Loop Engineering exploded. Addy Osmani said: 'You design the system that prompts the agent instead of prompting it yourself.'

A Loop is a 'volume.' Triggering, execution, observation, reasoning, verification, retry—all wrapped in a self-running, three-dimensional structure. Humans finally stepped back and became the designers of the cycle.

We were no longer satisfied with single deliveries and began pursuing continuous, autonomous, verifiable workflows. But when you only have one perfect volume in your hand, a new ceiling appears:

A single volume, no matter how fast, is still serial. Real business involves massive concurrency, requiring hundreds or thousands of volumes to run simultaneously.

The goal of a single volume is preset. But the value in complex systems often comes from the 'emergence' generated by the interaction of multiple volumes.

A single volume only verifies 'it's done,' not 'it's done right and is valuable.' If the direction is wrong, the faster it spins, the more efficient the error.

This is why after the Loop, it must be the 'Field.'

Field: From single-agent closed loop to field collaboration

If a Loop is the rotation of a single celestial body, then the 'Field' is a galaxy.

It no longer focuses on the perfect closed loop of a single agent but on the collaboration, evolution, and value alignment of multiple heterogeneous intelligent agents in time and space. This isn't a simple superposition of multiple agents but a leap across three dimensions:

01 From Master-Slave Relationships to Social Contracts

The Sub-agents in a Loop are hierarchical, whereas agents in a Field are peers. They communicate via protocols, negotiate tasks, and even engage in market-based bidding. Anthropic is exploring independent contexts and cross-validation between agents, while AWS emphasizes governance mechanisms at the agentic platform layer. This isn't gear meshing; it's a social division of labor.

02 From Task Completion to Value Alignment

A Loop pursues Proof-of-Done; a Field pursues Business ROI. Introducing a 'connoisseur mechanism' and a 'circular ledger' makes human feedback and business KPIs the gravitational field of the system. Agents are responsible not only for the code but also for costs, customer satisfaction, and revenue growth.

03 From Static Design to Adaptive Evolution

The rules of a Loop are set by humans; the structure of a Field is self-growing. The system can automatically adjust the number, division of labor, and skill combinations of agents based on historical performance. Inefficient Loops are eliminated, and efficient patterns are replicated. It's like biological evolution, not mechanical assembly.

Final Thoughts

From point to line, from plane to volume, and then to field. The essence of this path is the continuous upward shift of human control.

We no longer teach AI how to answer, no longer organize information for AI, no longer build cages for AI, and no longer even design AI's specific loops. What we begin to design is an ecosystem where countless loops spontaneously generate value.

In the next six months, when everyone has polished their Loops to near perfection, they will inevitably turn to the collaboration of the 'Field.' Because the era of going it alone is over. The next competition is about who can make a group of smart spheres dance in an orderly manner within the same gravitational field.

That's not a higher level of automation; that's the intelligent relations of production.

From point to line, plane to volume, and then to field, it's never achieved overnight, nor does it mean previous efforts are wasted. Rather, it's like building a skyscraper, one ring interlocking with the next.

After polishing one paradigm, the next paradigm is needed to solve more complex problems.

The term 'Field' hasn't yet been defined as the next keyword by any big name or paper. But it wasn't conjured out of thin air—it's the potential energy field in swarm intelligence, the bounded context in DDD, the agentic governance layer AWS is selling, a natural convergence on the line of AI engineering. We're just giving it a name, so that what is already happening can be discussed, designed, and replicated.