AI for the rest of us

Welcome to the AI community for everyone.

Hello friends,

Hannah is currently watching the rain from a lodge in Keldy Forest. A friendly squirrel just stopped by to say hello so she fed him a cherry tomato. Lucky squirrel!

This week Hannah has been in London at two very developer focussed conferences, AI DevCon and LeadDev (LDX3). There is so much change in the world of software engineering that instead of feeling empowered and informed, she left feeling more conflicted and confused than ever. Sessions contradicted each other with their advice and observations, people disagreed (amicably) and Hannah tried to filter through the hype and noise to find the actual signal. What is more clear than ever is that no one is coming to give us an easy answer. If you work in software today it’s time to roll up your sleeves and try something bold.

Charles was ignoring all of this, and spent Thursday at Lords cricket ground, for the first day of the 150th test to be played there. The forecast was dreadful, but despite interruptions to play he was treated to passages of outstanding bowling, with Kyle Jamieson claiming 5-62 for New Zealand and England’s Ollie Robinson marking his return with a compelling three-wicket maiden in his first over.

Have a great week!

Hannah & Charles

Agent Craft - This Week!

Join us in London for the pinnacle of London Tech Week. We're going deep on AI Agents in this one day unconference.

Our readers can grab £25 off the ticket price with code AI4NEWS25

What’s Charles reading this week?

I watched Casey Newton’s interview with Boris Cherny this week. Cherny is creator and head of the agentic coding tool Claude Code that Anthropic released last year and which is, by most measures, the fastest-growing AI coding tool in the world. He worked in a tiny team, and Claude Code was, rather like Gmail, a side project that took off.

The podcast is from a mini-series from Casey Newton’s Platformer about the impact of AI on jobs, and Cherny has said that he thinks software engineering could start to disappear this year, but it turns out he’s more optimistic than that implies. Whilst companies may hire fewer engineers as we know them today, he argues, they’ll hire more of whatever “builder” role replaces them. "I don't think we're going to call them engineers," Cherny says, "but if we talk about people writing code, or using agents to write code, I think there will be 100 times more engineers than there are today. That's my prediction.” My own view is similar: I think we’ll re-brand software engineering to something else, and I think the nature of the role will change — much like the move from punchcards to keyboards and screens — but I don’t think that the job itself goes away.

On Monday, Florida became the first state to sue OpenAI over ChatGPT’s allegedly dangerous design. Florida's civil suit is also aiming to hold Altman personally liable for alleged "reckless and wilful conduct", including showing "utter disregard for the risk to human life caused by his firms' conduct", according to the complaint. More than 20 lawsuits have already been filed against OpenAI over alleged ChatGPT harms, including by families of victims of a mass shooting at a school in Tumbler Ridge, Canada in February, and the families of seven people, including one teenager, who died by suicide or suffered delusions after using the chatbot.

I’ve been thinking about this a lot, but I’m not really settled on a view. I do think that OpenAI bears responsibility, but I am sceptical that the litigation route will produce good outcomes. Courts deciding case by case what AI companies owe their users is a blunt and slow instrument. What the situation probably calls for is something more like what we did with pharmaceuticals or aviation, i.e. a proper regulatory and testing regime before deployment, not liability claims after people are already dead. The problem though is that writing good legislation is slow and AI is moving extremely fast: and at the same time governments, at least in the US and the UK, are pushing AI in the hope that it will be a panacea for a lack of economic growth.

The Florida lawsuit is perhaps an attempt to force that reckoning through the courts in the absence of political will to legislate. Whether it succeeds legally is less important than whether it creates enough pressure to change behaviour, but I think that is also unlikely given the attitude of the people in charge of these companies which is, it seems to me, reckless.

I’ve not watched it for years, but I found myself thinking about Terry Gilliam’s Brazil whilst reading this article in 404 Media from Jason Koebler. Brazil describes a world where the future arrived but broke down immediately, and everything requires duct tape and repair. Koebler reports that hackers say that they used Meta’s AI support chatbot to break into a host of high-profile Instagram profiles by asking the support bot to change the email address associated with the target account. “The news shows the extreme risk associated with offloading support or critical functions to an AI chatbot”, Koebler says. “Users who have had their accounts stolen say that there is no way to escalate their problem to a human.”

I wouldn’t take it overly seriously, but I was amused by this report that Elon Musk’s chatbot Grok (or catbot as I initially wrote — I’m thinking the world needs one of them) oversaw complete societal collapse within just four days of being in charge of a simulated world. “The 15-day simulation saw Anthropic’s Claude establish a democracy with zero crime, with everybody surviving,” Anthony Cuthbertson writes in The Independent.

It is slightly off topic but I wanted to mention this podcast episode, because it is one of those stories I've wanted to do for ages and it has been really hard to get anyone to speak on the record. In my consulting work I'm having a number of conversations with people who are looking for options to move off US tech for obvious reasons, but an awful lot of options for "sovereign cloud" in Europe are rather nascent. Cycle.io is a US company, but they've recently set up a European control plane — fully segmented, running only on European-owned infrastructure providers, with no data or network overlap with their North American systems.

For the podcast I spoke to their co-founder and CEO Jake Warner, who has relocated to Iceland and is seeing a striking shift in demand: since last November, over 70% of new Cycle customers have included a bare metal component — often as part of a deliberate move away from the hyperscalers.

Finally, loyal reader Mark sent me this article from Micah Nathan, who teaches fiction writing at MIT, about students using AI. I don’t want to spoil it for you, because it is absolutely superb, but do read it.

What's Hannah reading this week?

Across the pond in San Francisco Microsoft hosted their massive annual conference, Microsoft Build. The big news that immediately hit my feed was MAI - Microsoft has been training their own frontier models (of course they have) and are ready to share them with the world. The tagline “Humanist Superintelligence” is both warm and terrifying. I don’t think the idea of an artificial superintelligence can ever be humanist? Can it?

Although most of my friends would flip a table if they were asked to use the Microsoft Office suite of products at work, there is no doubt that Microsoft can put those models to work immediately. Wherever OpenAI gets used today I presume? That can’t be good news for that partnership.

It’s almost impossible for Microsoft to fail at the AI game given they own everything from the productivity software, operating system and a global infrastructure of datacentres to train and run the models. They even have a ground breaking quantum computing chip Majorana 2 which is 1000 times more reliable than its predecessor.

This week in London I’ve been on a mission to find out what real teams are doing with agentic coding tools. What I found was a lot of conflicting experiences and opinions. There is the obvious friction between the AI Skeptics and AI Enthusiasts, which Charity Majors, CTO of Honeycomb so beautifully articulated this week in their piece “AI enthusiasts are in a race against time, AI skeptics are in a race against entropy”, but even amongst the AI enthusiasts there are massively diverging opinions about what good looks like.

I have so much to write this week I’ve had to try and collect my thoughts into three main categories.

Skills and Context Engineering

Back in the 2010’s it felt like we had to be talking about Cloud. It didn’t matter what the question was, or the problem at hand, the answer was CLOUD! Say cloud and you’re probably right. In 2025 the answer was MCP and Spec-Driven Development. Today in the software engineering world the answer is “Context Engineering and Skills” … it doesn’t really matter what the question is or the problem you’re solving.

It’s probably a moment in time. New buzz words will be arriving tomorrow.

I attended Patrick Debois’ talk where he explored the topic of Agent Enablement. Curating the context agentic coding tools need to do a good job: “We build the thing that builds the thing”.

That is certainly a new part of the job description of the software engineer but I don’t think it should replace the important job of “building the thing that our users need.” I spoke about this in my own talk (already available to watch here), proposing that some engineers will be naturally inclined toward more product shaped work solving users problems, and others will want to take on the more platform shaped work of building and improving the coding machine itself.

In discussion with Yenny Cheung, VP Engineering at Bluefish AI, she confirmed my suspicion that teams are starting to create new roles to bridge this gap. She’s hired an AI engineer to help Bluefish “Build the thing that builds the thing”.

I was surprised and delighted to have my talk delivered back to me as an agent skill. In fact all AI DevCon talks are now available as skills and all recordings are on youtube so if you missed the event you can benefit from all that goodness!

Ownership And Knowledge Sharing

Lawrence Jones, a Founding Engineer from Incident.io shared that right now ~20% of AI authored PRs are merged without changes, just a quick human check. He expects this number to continue to rise. That’s a pragmatic goal in my opinion. Rather than try and solve for every change, start to think about what changes you can make safely and start there.

There was also a human side to the lessons Lawrence shared at LeadDev. One of the strongest predictors of whether an engineer would embrace and enhance their skills in AI driven coding was whether they were working alongside someone who had gained some competence in the practices. Understanding that who you work with made a big difference the team at Incident started to rotate their most advanced engineers into other teams to spread the knowledge. So simple, no tech required.

Liz Fong-Jones (Technical Fellow at Honeycomb), Luke Marsden (CEO HelixML), Yenny Cheung (VP Engineering at Bluefish AI) and I had a fabulous debate on the topic of code review. I don’t believe mandatory code review is sustainable and we need to find other ways to serve the purpose. Liz countered that building software products is a team endeavour and that reviewing each other’s work is an important part of knowledge sharing. That is also true. This is one of the big areas of conflict in software engineering teams today.

I respect Liz enormously, Liz is far smarter than me, but I think we need to find a new way to share knowledge that doesn’t involve reviewing hundreds of lines of AI authored code. Spec-reviews could be the answer or tried and tested practices like pair programming! Reading code is not it. Not anymore.

I also loved this piece of writing on notyourpeer.com about the importance of helping each other, passing on our experience, asking for help and sharing our knowledge.

“AI models lack the things we mean when we say peer: passion, inspiration, aspiration, the ability to invent. They don't bring a point of view. They don't disagree with us because they think we're wrong. They are competent, patient, tireless, and none of that is what makes a peer worth having.”

In her closing keynote at AI DevCon, Birgitta Boeckeler (Distinguished Engineer at Thoughtworks) talked about her current discomfort, asking “what are we surrendering to AI?” and urged teams to be intentional and mindful about the things they are ready to let go and what they want to keep no matter what … It was an enormous privilege to be quoted by Birgitta, who referred to my keeping / trashing / trying framework.

Velocity and Parallelisation

For the first time ever I heard an engineering leader (Chris Batey CTO of Core Engineering Consulting Group) propose that his engineers take on less work. LESS WORK! What!?

The temptation to kick off another agent and parallelise work is massive, especially when you often have to wait a while for an agent to return a response. However, Chris had observed that the quality of the work was degrading as his team picked up more work in parallel. I loved that Chris has extended the role of each engineer to include “features adopted by users” to ensure they are always improving their products and not just shipping features. It’s a 30 minute talk full of real world experience, you can watch it here.

Kanban is a process optimisation practice that is popular in software development. I love it. Kanban helps improve flow throughout a whole end to end process by strictly limiting the work in progress. Don’t add something new until the current thing is done. Avoid distractions and context switching. It feels like many teams are losing sight of that as they drive their teams towards more and more parallelisation of work.

Or perhaps, as Luke Marsden commented, we need engineers who can context switch across multiple tasks the way engineering managers do today. There is no clear answer here, both seem like reasonable opinions about what the future might look like.

In discussion with an engineering leader from Grafana at LeadDev we also hit upon the same thing. Flow. They have moved to 100% AI driven SDLC with minimal code review or human intervention. They achieved this by breaking down agentic coding tasks into small low risk changes. I like this because the engineer is making all the decisions up front. It’s highly controlled but it’s still fast, especially if that code can reach production without any further human intervention.

The team at HelixML do parallelisation slightly differently. They have a centralised agent platform to allow teams to manage agents collectively. It’s not 1 developer to many agents, it’s many developers to many agents. I had seen it in action before but this week I suddenly got it. If a team collectively manages multiple agents the work that the agents are doing becomes transparent to the whole team, everyone can review specs and track progress. Everyone has context (there’s that word again) about the changes being worked on and maintain a shared ownership of that code in production. Luke explained his reasoning and the benefits of a shared agentic platform in his talk on Monday.

The Confusion and Conflict

Experimentation is good and I can see so many teams throwing out the old and trying something radical and new. Some of the things that made human teams successful will inevitable slow us down in our new world of Agentic Software Development. We must be wise about what we hold on to and what we throw away.

My observation from this week is that teams are tackling the same challenge in very different ways, making very different decisions about what the things they choose to preserve. It will be fascinating to see what works in the long run.