Each engineer I do know is coding with AI now. Brokers are writing an enormous quantity of code, and but someway no one is transport 5x extra software program. Groups obtained their 30% bump when AI coding mainstreamed after which that productiveness flatlined. If this resonates or sounds acquainted, it’s as a result of we have now to cease enthusiastic about learn how to immediate brokers and begin enthusiastic about the programs round them.
You could suppose you’re already a part of the “new” engineering regime should you’re constructing with brokers. Nonetheless, prompting an agent to code was solely state-of-the-art in 2025; now we’re half-way via 2026 and the brand new “new” engineering is constructing the system that builds the system. For the groups which have embraced the software program manufacturing unit, the good points are compounding to 3x and past however for everybody else the AI-gains are stalled.
Because the famous writer William Gibson famously mentioned, “The longer term is already right here, it’s simply not evenly distributed.”
There’s a large consideration shift wanted if you wish to transcend the 30% productiveness achieve and get to 3x utopia. As a substitute of enthusiastic about how you’re constructing, shift to what you’re constructing. Prompting an agent to code is simply constructing a quicker horse; designing a system of brokers is constructing a automotive.
There are three facets about LLMs that turned true on the finish of 2025 that allow this transformation, and that is what’s required to embrace the “new new engineering” mindset:
- The flexibility for LLMs to finish longer and extra complicated duties is accelerating.
- LLMs are able to producing the suitable output, so long as they’re given the suitable inputs.
- Brokers have develop into self-driving.
1. Process functionality is accelerating
Contemplating that LLMs are able to doing extra and that price is simply rising, then embracing the “new new engineering” mindset is required to alter the scope of labor you’re handing to them. This implies:
- Go larger: take into consideration work you possibly can hand off that takes hours for the agent to finish, not minutes.
- Get off your laptop computer: an agent that’ll run for hours can’t be sure by your community connection.
- Get out of the best way: your consideration is the largest interference to an agent working for hours.
2. Output is sure by enter
The flexibility for an LLM to provide the suitable output is now strongly correlated to the standard of the enter. Inputs go nicely past prompts and embrace plans, specs, examples, codebase, and suggestions loops. In case your brokers can’t run greater than 5 minutes at a time with out going sideways, that’s a symptom of lacking context.
Suggestions loops are particularly highly effective as they permit brokers to confirm their work. The brand new mindset requires a shift that brokers can construct their very own instruments to self-verify. Check suites and browsers are simply the beginning of verification, brokers can construct their very own instruments to self-verify. These instruments may be brief lived for a particular activity or undertaking, or lengthy lived like a reference implementation of an SDK; each of which had been impractical earlier than AI however at the moment are trivial.
3. Brokers are self-driving
Brokers needs to be doing a lot of the driving, you’re designing the roads and giving them the routes. Greater than self-driving, brokers needs to be spinning up different brokers. A contemporary coding loop incorporates brokers performing as coordinator, planner, implementer, verifier, and reviewer.
So as to construct an working system of brokers:
- Do one factor: construct single-purpose, centered brokers and expertise that may be chained collectively to perform duties.
- Adversarial brokers: design brokers that require proof earlier than permitting the workflow to proceed.
- Self-improvement: when one thing fails, use brokers to enhance the system for the subsequent time.
The brand new new engineering is just like the outdated engineering
This may occasionally appear daunting and have you ever questioning your expertise; however the “new new” engineering isn’t that far off from what we’re used to. The software program engineering artwork is alive and thriving, it simply requires a mindshift together with:
- Software program engineering was all the time about constructing a system of inputs and outputs; code was simply the medium.
- You don’t need to accept spaghetti code and AI slop; should you don’t just like the output then enhance the enter.
- Your experience is extra essential than ever. When you didn’t know what good code is meant to seem like within the first place, then how will you design a system to provide good code?
- You’re nonetheless transport software program as a result of you possibly can’t design the system with out having actual work to get executed. It’s a must to construct the aircraft in flight.
Altering how code will get generated will get us the 30% good points, however there’s a complete system of software program growth that must be rethought to get to the 3x promised land.
Simply do not forget that we’re solely getting began, and there’s doubtless a “new, new, new engineering” coming in 2027.
