On the GOTO Convention in Copenhagen in 2025, Kent Beck and I spent a while on stage speaking and answering questions from the viewers – a format I confer with as “two outdated geezers on a park bench”. We discuss our experiences with LLM-augmented programming (at that time – October 2025), we present our frustration that issues we’ve been saying for thirty years nonetheless should be stated, we are saying how something like a manifesto reunion must be led by a youthful technology, and opine on what junior builders must be specializing in of their profession.
❄ ❄ ❄ ❄ ❄
Ian Johnson has written a sequence of posts about restructuring a gnarly codebase
The story follows an actual Laravel + React codebase over ~3 months and ~258 commits from a legacy monolith with no exams to a well-structured utility with automated high quality gates, a React SPA migration in progress, and an AI agent that reliably ships manufacturing code with minimal supervision.
The sequence covers the steps in respectable element, and his strategy follows the sorts of steps I’d use. First get every little thing beneath the management of respectable characterization exams, add static evaluation, introduce the proper patterns to make issues move simply.
With all of this, is his use of AI, which modified in the course of the train:
For the primary two months of this undertaking, I used Claude Code with auto-approve turned off. Each file edit, each terminal command, each change… I reviewed it earlier than it executed. […] The outcomes have been good. The code was clear. However I used to be doing many of the considering and half the typing. The agent was a elaborate autocomplete with higher ideas. I wasn’t getting the leverage I’d hoped for.
I learn an article about “on-the-loop” versus “in-the-loop” human-AI collaboration. The framing clicked instantly […] I used to be micromanaging as a result of I didn’t belief the agent to do the proper factor. And I didn’t belief the agent as a result of there was nothing forcing it to do the proper factor.
His early steps put in exams, static evaluation, and the proper architectural patterns. With these in place, he may let the agent do extra work.
My function shifted from author to curator. I don’t write many of the code anymore. I Outline the patterns […] Evaluate the check specs […] Evaluate the output […] Replace the harness […] Make strategic selections […]
He finishes the sequence with conclusions about how he’d generalize his expertise to different circumstances.
❄ ❄ ❄ ❄ ❄
Again within the land of my start, there was some notable groans when the Nationwide Well being Service determined to shut practically all of their Open Supply repositories, supposedly to the safety menace of LLMs. Closing repos like this isn’t an efficient counter to LLM-augmented attackers. I think it’s no coincidence to see GDS (Authorities Knowledge Providers), the highly-regarded IT enablers within the UK authorities publish their place
Transferring code from public to personal as an alternative choice to funding in secure-by-design supply, possession and remediation is a warning signal as a result of it reduces sharing and scrutiny, can gradual coordinated enchancment throughout authorities and suppliers, and doesn’t take away the underlying weaknesses in a operating service.
Terence Eden memorably sums up his view on this:
Inside the UK’s Civil Service you often hear the expression “being invited to a gathering with out biscuits”. It implies a somewhat frosty dialogue with none of the well mannered niceties of a traditional assembly.
❄ ❄ ❄ ❄ ❄
I’ve seen a couple of circumstances the place these builders who’re most concerned in working with LLMs discover they’re operating into an issue with cognitive endurance, Adam Tornhill has joined this group:
One of many huge wins with brokers is that they allow us to stick with the higher-level downside for longer. We get much less sidetracked by particulars, dependency cleanup, and related secondary duties that used to interrupt focus.
However there’s a price we’re nonetheless underestimating. Agentic coding is mentally costly.
I can often maintain the tempo for a few hours. Then I would like a break. The tempo is just too intense. And based mostly on conversations with different engineers, I don’t suppose I’m alone in that.
He explains that working with The Genie means we’re making extra selections in much less time, this improve in determination density is difficult on the mind.
He responds by retaining agent duties small, automating every little thing he can, and accepting that he received’t know each line of code so long as he has good verification mechanisms in place.
Notably, he has not gone within the path of doing his work with swarms of brokers that he coordinates. As a substitute has one long-running process that he babysits and one focus process
That final level is essential given the running-twenty-agents-in-parallel hype. I can not even take into consideration twenty significant issues to construct, and even much less so concerning the ensuing cognitive tax of the probably interruptions. It’s precisely the flawed factor to even think about. No less than for people. (And sure, I perceive sub-agents and machine parallelisation. That’s not what I’m objecting to. It’s the parallelisation of human consideration that doesn’t scale).
I appreciated that he included some ideas about what of us can do in time outdoors this intense programming time. Not simply “have a espresso” (though he consists of that) but in addition about studying concerning the area that the software program helps.
❄ ❄ ❄ ❄ ❄
A few pithy quotes from social media
“Metaphor debt” is when your whole metaphors contain the idea of “debt” as a result of you possibly can’t consider some other metaphors anymore.
❄ ❄
If a vegan crossfit fan is utilizing Claude to jot down Rust, which factor do they inform you first?
❄ ❄ ❄ ❄ ❄
Karl Bode reacts to audio system getting booed when mentioning AI throughout graduation addresses. He factors out that youthful of us are more and more sad with the tech oligarchy and their fruits.
The factor is the children aren’t silly. They see the sphere clearly. They see the distinction between what’s being offered to them by tech corporations, the press, and graduation audio system, and what they’ve repeatedly seen with their very own eyes.
They’ve watched tech oligarchs spend the final decade mired in scandal after scandal, hype cycle after hype cycle, steadily enshittifying every little thing they contact alongside the way in which.
[…]
The share of Gen Z that suppose AI’s advantages don’t counterbalance the dangers now sits round fifty %, up 11 proportion factors in simply the final 12 months. Eight out of each ten consider that utilizing AI makes the method of precise studying harder.
He sees younger individuals saddled with the notion of getting into a worsening world –
which leads them to rage in opposition to this newest fruit of the tech oligarchy. A rage
that’s straightforward for folk like me
– with a snug retirement off-ramp – to correctly admire. A rage that would have marked political and social penalties.
❄ ❄ ❄ ❄ ❄
Related to those considerations are a few gadgets in final week’s Economist newspaper. The newspaper argues that traditionally main technological advances haven’t led to important unemployment or drops in wages (paywalled article). The closest was the unique industrial revolution in nineteenth Century Britain. There was a stagnation in wages throughout this era, however there was additionally an enormous improve in inhabitants, from 4½ million to 12 million.
It additionally factors out that we’ll most likely solely perceive the complete penalties of all this when a recession hits, as that is when most unproductive jobs are typically flushed out of the system.
A second article (additionally paywalled) signifies that AI is having some impact on graduate hiring. They did an evaluation of surveys of latest graduates, trying to see if employment assorted relying on a job’s publicity to AI. The least uncovered quintile of topics noticed employment charge fall by 1.5% over the past couple of years, whereas probably the most uncovered quintile’s drop was 6.6%.
❄ ❄ ❄ ❄ ❄
Lawfare isn’t impressed with the most recent efforts by the US Authorities to control AI.
On [last] Wednesday, the White Home invited leaders of OpenAI, Google, Anthropic, Meta, and Microsoft to the Oval Workplace for a signing ceremony the next afternoon. President Trump was to signal an government order on AI and cybersecurity—the administration’s most formal effort but to determine a voluntary course of for reviewing frontier fashions earlier than their launch. However roughly three hours earlier than the ceremony, when some firm executives have been already within the air to Washington, the White Home known as it off.
They see the proposed rules as gentle, and together with some helpful measures to harden defenses in opposition to cyber threats.
But it surely’s price underscoring the implications of suspending (if not outright canceling) this order, which, by its personal phrases, was about as modest a frontier-AI intervention because the federal authorities may placed on paper: voluntary, targeted on the federal government’s personal defenses, and explicitly barred from changing into a licensing regime. The objection isn’t a lot about authorities coercion as concerning the authorities having any settled function in any respect. Voluntary, in different phrases, isn’t the ground of frontier AI coverage on this administration; it’s the ceiling.
This can be a questionable place provided that the considerations animating this draft order will probably develop within the close to future. It is usually self-defeating for individuals who applauded the order’s delay or demise. Removed from resolving the chance of presidency meddling in AI, killing the order simply leaves in place what Ball has described because the “opaque and basically lawless” different: authorities entry occurring by way of again channels, on phrases set case by case, with no secure guidelines in any respect.
One of many issues here’s a distinct lack of governmental experience, both in AI or in software program basically. An excessive amount of is being determined on the whims of the tech oligarchy, there isn’t any try to have interaction within the broader points at hand. That’s not fully a nasty factor, attempting to control one thing that’s nonetheless evolving so quick is often a idiot’s errand – however the issue right here is the influence of AI is so huge that there’s actual hazard in being too far behind.
❄ ❄
Which leads me to a uncommon factor, an endorsement of a candidate for political workplace. If you’re voting in congressional district MA-06 (North Shore of Massachusetts), I’d significantly have a look at Beth Anders-Beck, who’s operating for congress in that district. Beth has an extended background in software program improvement (together with creating the notion of Forest and Desert), so would introduce experience that Congress desperately wants. I’ve recognized Beth for many years, and have a excessive opinion of their intelligence, judgment, and talent to work with others. Congress doesn’t deserve Beth, but it surely does want her.
