Table of Contents
Every engineer I know is coding with AI now. Agents are writing a massive amount of code, and yet somehow nobody is shipping 5x more software. Teams got their 30% bump when AI coding mainstreamed and then that productivity flatlined. If this resonates or sounds familiar, it’s because we have to stop thinking about how to prompt agents and start thinking about the systems around them.
You may think you’re already part of the “new” engineering regime if you’re building with agents. However, prompting an agent to code was only state of the art in 2025; now we’re half-way through 2026 and the new “new” engineering is building the system that builds the system. For the teams that have embraced the software factory, the gains are compounding to 3x and beyond but for everyone else the AI-gains are stalled.
As the noted author William Gibson famously said, “The future is already here, it’s just not evenly distributed.”
There is a massive attention shift needed if you want to go beyond the 30% productivity gain and get to 3x utopia. Instead of thinking about how you’re building, shift to what you’re building. Prompting an agent to code is just building a faster horse; designing a system of agents is building a car.
There are three aspects about LLMs that became true at the end of 2025 that enable this transformation, and this is what is required to embrace the “new new engineering” mindset:
- The ability for LLMs to complete longer and more complex tasks is accelerating.
- LLMs are capable of generating the right output, as long as they’re given the right inputs.
- Agents have become self-driving.
1. Task capability is accelerating
Considering that LLMs are capable of doing more and that rate is only increasing, then embracing the “new new engineering” mindset is needed to change the scope of work you’re handing to them. This means:
- Go bigger: think about work you can hand off that takes hours for the agent to complete, not minutes.
- Get off your laptop: an agent that’ll run for hours can’t be bound by your network connection.
- Get out of the way: your attention is the biggest interference to an agent running for hours.
2. Output is bound by input
The ability for an LLM to produce the right output is now strongly correlated to the quality of the input. Inputs go well beyond prompts and include plans, specs, examples, codebase, and feedback loops. If your agents can’t run more than 5 minutes at a time without going sideways, that’s a symptom of missing context.
Feedback loops are especially powerful as they allow agents to verify their work. The new mindset requires a shift that agents can build their own tools to self-verify. Test suites and browsers are just the start of verification, agents can build their own tools to self-verify. These tools can be short lived for a specific task or project, or long lived like a reference implementation of an SDK; both of which were impractical before AI but are now trivial.
3. Agents are self-driving
Agents should be doing most of the driving, you’re designing the roads and giving them the routes. More than self-driving, agents should be spinning up other agents. A modern coding loop incorporates agents acting as coordinator, planner, implementer, verifier, and reviewer.
In order to build an operating system of agents:
- Do one thing: build single-purpose, focused agents and skills that can be chained together to accomplish tasks.
- Adversarial agents: design agents that require proof before allowing the workflow to proceed.
- Self-improvement: when something fails, use agents to improve the system for the next time.
The new new engineering is similar to the old engineering
This may seem daunting and have you questioning your skills; but the “new new” engineering is not that far off from what we’re used to. The software engineering art is alive and thriving, it just requires a mindshift including:
- Software engineering was always about building a system of inputs and outputs; code was just the medium.
- You don’t have to settle for spaghetti code and AI slop; if you don’t like the output then improve the input.
- Your expertise is more important than ever. If you didn’t know what good code is supposed to look like in the first place, then how can you design a system to produce good code?
- You’re still shipping software because you can’t design the system without having real work to get done. You have to build the plane in flight.
Changing how code gets generated gets us the 30% gains, but there’s an entire system of software development that needs to be rethought to get to the 3x promised land.
Just remember that we’re only getting started, and there is likely a “new, new, new engineering” coming in 2027.
