I’ve spent a lot of my profession as each an analyst and a practitioner advising folks towards “the massive rewrite.” It’s referred to as second-system syndrome for a motive: You often throw away years of hidden bug fixes and safety patches only for the sake of “cleaner” code. However after sitting down with Cloudflare’s Dane Knecht (CTO) and Steve Faulkner (engineering director) to debate their challenge vinext, it’s clear that the previous guidelines of technical debt are being rewritten by agentic software program improvement.
Actuality Examine: The Finish Of Monkey Patching
For the final 18 months, the Cloudflare neighborhood relied on OpenNext to run Subsequent.js on their infrastructure. It labored, however as Dane famous, it wasn’t a first-class expertise. “It was a hack, you recognize, in a way … the place you’re sort of monkey-patching it to attempt to work in different places.”
In a conventional software program improvement lifecycle, you reside with that technical debt as a result of a rewrite is simply too costly or impractical. However Cloudflare took a special path. By transferring to a very agentic workflow (the place, as Dane put it, “no [human] palms have been on the keyboard”), they rebuilt the framework from first rules.
The consequence? A codebase that’s smaller, quicker, and optimized for a world the place developer expertise typically takes a again seat to agent expertise (AX).
A New Rhythm: The three A.M. Barbell
The world (myself included) has develop into obsessive about how AI adjustments the rhythm of labor. Steve shared one thing that ought to make each engineering chief pause: His workflow has develop into “barbell-shaped.”
“Both I had very brief like 5–10-minute classes focused at an issue or these like 5–6-hour-long classes,” Steve defined. However the actual shocker was when he checked his session information: He observed his peak token utilization was usually at 3 a.m.
“I’m not awake at 3 a.m.,” Steve informed me. “I’d simply brain-dump … right here’s 5 issues to go work on whereas I’m asleep and, like, see how far it will get.”
That is the shift from “supervisor” to “dispatcher.” Steve emphasised that he isn’t a “100x engineer” — he’s an engineering supervisor who has realized to handle brokers in parallel whereas he sleeps.
The “Mirror Downside”: AI Inherits Our Human Challenges
An fascinating second of the challenge was a safety oversight that completely illustrates a core problem with agentic improvement. Throughout the preliminary rewrite, the AI brokers repeated the very same safety vulnerabilities (CVEs) discovered within the unique Subsequent.js supply code.
“AI is extra humanlike than you anticipate in quite a lot of methods,” Dane noticed. “It made the identical errors that they made once they had been writing it initially.”
He famous that this one change to their prompting would have significantly improved the challenge’s safety profile: “Please assessment all previous CVEs on Subsequent.js and make sure that we’re not susceptible.” It was a lesson in context. Whereas the AI missed vulnerabilities at first as a result of it was mimicking its coaching information, Dane famous that when they pointed it to the fitting corpus of knowledge, it “instantly recognized them and was in a position to repair them.” The takeaway for the maturity curve is that agentic outcomes are a mirrored image of the context you present. Should you don’t explicitly take motion to keep away from your previous errors, brokers will recreate them with excellent, automated effectivity.
The Maturity Curve: Biased Towards The Merge
Maybe probably the most fascinating takeaway is Steve’s philosophy on merging code. Many organizations are gun-shy about merging, as a result of fixing manufacturing is pricey.
On this case, Cloudflare is doing the other. “That’s been our default … we’re sort of, like, biasing towards merging,” Steve mentioned. “We simply know that there are bugs. We all know there’s gaps. We all know there’s stuff that’s received to get mounted. So we’re simply trusting that the neighborhood comes up with good options.”
This works as a result of the price of an agentic repair is close to zero. If a bug is discovered, an agent triages it, writes the repair, and submits a pull request. Are we reaching a stage of maturity the place the velocity of correction is extra worthwhile than the knowledge of the commit?
Put together For AX
Steve made a remark that can follow me for a very long time: The vinext repository is definitely troublesome for a human to navigate. “Should you’re not interacting with this repo with AI, you’re going to have a nasty time,” he warned. “It’s truly actually not designed for you. It’s designed for AI-first approaches.”
This can be a key a part of AX: As we transfer additional up the maturity curve, we should begin constructing environments the place brokers can thrive.
A lot of the chatter on this area is about how AI could make builders extra productive. The trail to agentic ROI begins with the belief that sunk price and technical debt imply one thing completely different at this time than it did previously — particularly in the event you’re courageous sufficient to let the brokers take the keyboard.
