Every developer I know is coding with AI now. Agents write a huge amount of code, and yet somehow no one ships 5x the software. Teams get their 30% bump when the AI coding is overplayed and that produces a fade. If this sounds or sounds familiar, it’s because we have to stop thinking about how to inform agents and start thinking about the systems around them.
You might think you’re already part of the “new” engineering realm if you build with agents. However, encouraging an agent to code was only a state of the art in 2025; now we are in the middle of 2026 and the “new” engineering is building a system that is building a system. For teams that have adopted a software factory, the benefits compound to 3x and more but for everyone the benefits of AI are stagnant.
As noted author William Gibson famously said, “The future is here, it’s not evenly distributed.”
There is a major shift in focus required if you want to surpass the 30% productivity gain and get to 3x utopia. Instead of thinking How you create, switch to what you build. Telling the agent to enter the code is just to create a fast horse; designing a system of agents to build a car.
There are three aspects about LLMs that will become reality by the end of 2025 that allow this change, and this is what is needed to embrace the concept of “new engineering”:
- The ability of LLMs to complete long and complex projects is increasing rapidly.
- LLMs are capable of producing the right output, provided they are given the right input.
- The agents are already driving.
1. Labor force is increasing rapidly
Considering that LLMs are able to do more and that level is only increasing, then embracing the concept of “new engineering” is necessary to change the scope of the work you offer. This means:
- Grow: think of a job you can offer that takes hours for an agent to complete, not minutes.
- Remove your laptop: An agent that will work for many hours cannot be bound by your network connection.
- Get out of the way: Your attention is a huge distraction for an agent working hours.
2. The output is tied to the input
LLM’s ability to produce correct output is now closely related to the quality of the input. Input goes beyond information and includes programs, specifications, examples, codebase, and feedback loops. If your agents can’t work for more than 5 minutes at a time without going to the side, that’s a sign of a lack of context.
Feedback loops are particularly powerful as they allow agents to validate their work. A new way of thinking requires a change where agents can create their own tools to validate themselves. Checkpoints and browsers are the starting point for authentication, agents can build their own tools to authenticate themselves. These tools can be short-lived for a specific task or project, or long-lived as a reference SDK implementation; both of which were useless before AI but are now useless.
3. Agents drive themselves
Agents should be driving a lot, designing roads and giving them routes. In addition to driving themselves, agents must drive other agents. A modern code loop includes agents that act as a coordinator, organizer, initiator, verifier, and reviewer.
To create an agent application:
- Do one thing: create single-purpose, focused agents with skills that can be tied together to accomplish tasks.
- Adversarial agents: design agents require proof before allowing the workflow to continue.
- Self-improvement: if something fails, use agents to improve the system in the future.
New engineering is like old engineering
This may seem difficult and question your abilities; but the “new” engineering is not so far from what we are used to. The art of software engineering is alive and well, it just needs a mindset shift that includes:
- Software engineering has always been about creating system inputs and outputs; the code was just in.
- You don’t need to find spaghetti code and AI slop; if you don’t like the output improve the input.
- Your technology is more important than ever. If you didn’t know what good code should look like in the first place, how could you design a system to generate good code?
- You still ship software because you can’t design a system without having real work to do. You have to build a plane to fly.
Changing the way code is generated gives us a 30% gain, but there is an entire software development process that needs to be rethought to get to the promised land of 3x.
Just remember that we’re just getting started, and there may be “new, new, new engineering” coming in 2027.
