Code doesn't have to ship
Coding agents make more engineering questions cheap enough to test.
In April, a conversation with 
Tanner Linsley at AI Engineer Miami and React Miami left me with a question: what if we used coding models to test new shapes for a product, not just clear its existing backlog?
I had somewhere to test it. For c15t v3, I was working on the server-rendered path for consent experiences.
Client-side c15t has a straightforward request path:
Client → c15t backend → Client
Server-side rendering lets c15t use the framework to prepare the geo- and locale-specific experience before the page reaches the browser. In Next.js, that path looked like this:
Client → Next.js server → c15t backend → Next.js server → Client
This moved the work server-side, but the Next.js server still had to call the c15t backend before it could respond. That put the latency between those servers on the critical path.
I wanted to know whether the manifest system I built for scripts in v2 could also describe the data needed to render a consent experience. My hypothesis was that we could cache that manifest at the CDN edge, then let the Next.js server combine it with the request's geography and locale to produce the response itself:
Client → Next.js server → Client
That would keep the benefits of server-side rendering while removing the trip to the c15t backend. The question was whether the performance improvement would justify moving more logic into the manifest system.
Proving it by hand would have meant giving an uncertain direction real roadmap time. That is usually where an experiment loses to work with a clearer payoff.
Instead, I wrote down the hypothesis, the rough implementation, and the benchmark that would decide whether it was worth pursuing. I set an agent loop running, then spent the day cleaning my apartment.
The experiment worked. With the manifest cached at the CDN edge, the Next.js server could produce the response without waiting for the c15t backend, and the benchmark improved enough to keep going.
Then I tested it in real production environments. Latency came back, and server-side rendering had problems the benchmark never showed. I was ready to scrap the idea.
Before I did, I gave another agent a narrower question. What if the build generated the manifest and shipped it inside the bundle? With nothing to fetch at request time, the Next.js server has nothing left to wait on. That fixed the main problem, and it's the default setup in c15t v3. In my latest benchmark, the banner showed in 27ms on v3, down from 336ms on 2.3.
The edge-cached version didn't become the default. It still ships as a fallback for policy changes that can't wait for a rebuild. It also showed that removing the backend trip mattered, which told the second experiment where to look.
I started using the same approach elsewhere in c15t and across a few of my other projects. Some ideas worked. Others failed miserably.
The failures were useful. While a model explored a dead end, I worked on something else. I still had to inspect the implementation and decide what its result proved, but I did not have to build every branch myself.
Theo described the distinction: code can be useful without being merged.
Theo - t3.gg@theoThat is what makes this different from asking a coding agent to clear a backlog. The task is not “implement this feature.” It is “produce enough evidence to decide whether this idea deserves to become a feature.”
A prototype gives an idea shape. An experiment answers a question. Coding agents make both cheaper, but only the second starts with a hypothesis and a way to prove it wrong.
More possible implementations do not create more time to evaluate them. Every branch still asks for attention. Guillermo Rauch makes the other half of the argument: greater implementation capacity makes engineering judgment more important, not less.
Guillermo Rauch@rauchgThat is why I start with a question and a way to measure the answer. Without both, cheap code just creates more code to review.
The code does not need to ship. It just needs to make the next decision clearer.