This focus on maintenance is a really interesting approach, and has a lot of merit as we move into an era where a significant amount of AI generated code will need to be sustained as part of production systems. Cognition get a lot of extra points in my mind for including Java and C/C++ in their evaluation. Not everything is greenfield or a rewrite from scratch.

Cognition’s blog follows interesting work that Cursor dropped a few months ago with the publication of Cursorbench and DataCurve released DeepSWE.

Now all of these evals are trying to address gaps which existing in the widely used SWEBench family. SWEBench has encountered the age old problem of all benchmarks - people will lay train tracks through their tooling to improve marketing metrics.

That said, SWE-bench Pro is very widely used, and provides a little glimpse into how everything is converging. I keep an ongoing chart of results and trends across major models, which I update every month or so.The trend is pretty clear, for many basic tasks we are fast hitting a convergence point. We are well past the point where you need the latest greatest model for a lot of day to day work.

The latest piece, thoughtfully crafted in under 30 seconds by Anthropic’s Claude Opus 4.8” is Lunch is a Legacy System: Deprecating Human Needs for Q4 Targets. Abstract below. Enjoy..

“Most enterprises rigorously audit their software estate while ignoring the largest unmanaged runtime they own: their people. This Point of View introduces Necessity Rationalization™, a four-stage model for treating inherited human needs—starting with lunch—as legacy systems flagged for deprecation.”

This acquisition fits in with a trend of developer tool acquisitions over the last eighteen months by Cloudflare, Vercel, OpenAI and Anthropic. Broadly these fit into two categories:

• infrastructure tooling for language ecosystems, used by agentic coding agents (OpenAI, Anthropic and other consumers)
• front end javascript ecosystem tooling (Cloudflare, Vercel)

with the outputs of the coding agents supporting deployments on vendors such as Vercel and Cloudflare.

The notable exception here is Stainless, who provided tooling for generating SDKs and so forth, and were recently acquired by Anthropic. Stainless was also used by Anthropic competitors, and the acquisition is both an acquihire and a (short term) speed bump for their competitors.

The focus on JavaScript/TypeScript and Python tooling makes perfect sense. The growth of TypeScript and Python was highlighted in the 2025 GitHub Ocotoverse report. These languages are, by far, the preferred choice of most LLMs. The growth in the use of both languages has only accelerated since the November 2025 model inflection point.

Developer Tool Ecosystem Acquisitions

Anthropic, OpenAI, Vercel, Cloudflare since July 2025. Not a complete list.

• Sandboxes for Copilot. This brings simple, clean, isolation when your using the CLI and other form factors (I live in the CLI these days, although I am increasingly using the new GitHub Copilot App). Longer term sandboxes are going to be the only way enterprises will permit agents to run either locally or in the cloud.

• MAI-Code-1-Flash Model. MAI-Code-1-Flash is a new, Microsoft developed, model. It excels in price/performance for light weight coding flows. How we use models is rapidly changing, and the days of rushing to the latest shiny (and uber expensive) frontier model are coming to a close.

• Learning from your sessions with /chronicle. Learning from, and improving, how we use AI - what we do in our sessions - is going to be a key skill for every developer. Your sessions across surfaces can be synced to GitHub, and using /chronicle you can do things such as creating summaries of what you have done, suggesting improvements and working out more cost effective ways of working. Longer term the possibilities here are endless.

I have found this entire space fascinating since 2008, and I wanted a simple way to plot and compare the overall growth over time. So I created a Cloud Region Launch History tool. You can filter to specific times, across providers, sovereign regions, government clouds and industry specific offerings.

For example my analyst career started in 2015 at RedMonk, the map for AWS and GCP looked a bit like this:

And by the time I led the Cloud Application Platforms research at Gartner in early 2021 the map was:

When I left the analyst game and joined GitHub later in 2022, Azure’s footprint was:

And these days it looks like:

The tool is available to anyone to use. Constructive comments and feedback on any obvious glaring data gaps (there are a few where the number of AZs needs to be further verified) appreciated.

Notes on the Data

Data is sourced from public announcements and documentation. We will have gaps where vendor specific details are behind a login or paywall. We include a filter to eliminate regions with less than 3 AZs. That said, the definition of an AZ, and in particular the geographic distances between each AZ, varies between providers - so do check this data yourself.

Providers covered:

We included the three hyperscalers:

• AWS
• Azure
• GCP

And a set of smaller or regional providers:

• Alibaba
• Huawei
• IBM
• Oracle
• TenCent

The providers I included for this iteration are the same as those Gartner included in the (catchily titled) Magic Quadrant for Strategic Cloud Platform Services, published in August 2025. Gartner have a minimum revenue requirement of at $1Bn for vendors with more than three years GA, or $500M with 40% CAGR for those who have been GA for less than three years. There are a bunch of other criteria, but I find this useful as a guideline.

Disclosures: I covered cloud, and specifically application development in cloud environments extensively at Gartner and RedMonk. During that time I attended numerous analyst events related to each of the hyperscalers. I currently work at GitHub.

I covered modernization (thankfully not Cobol specifically though) a lot when I was at Gartner. I have had more briefings and demos for magic bullet solutions than I care to remember. Including the magic bullets IBM were developing in IBM Labs that became part of the catch all watsonx portfolio, albeit IBM were working from a far more defensive posture ².

The difficulty is not, and has never been, just the code - it is everything around it. The people, processes, systems, reliability requirements, compliance and so forth all come into play. I have no doubt at all that AI can and will hugely accelerate some of the technical side of the changes - if there is an appetite in the org do to do it. I already see this a lot in my day-to-day work.

However, the attitude of the organisation – the digital literacy – is probably the most important factor. How do they view the ongoing maintenance of their investments in technology. Are technology projects one off investments – get me from point a to point b – or has there ever been an appetite for ongoing investment. Care and feeding. Ask yourself the question how did the organisations that are desperate for this magic bullet end up in this mess in the first place?

This pattern, that of people, processes, systems, reliability and compliance is the same pattern I often discuss when talking about DevOps ³. Leading organizations are way out in front, but people don’t realise just how far behind many companies really are. AI can be, and is, a massive accelerator here as well, once there is appetite for the level of change required and the skills to implement it.

The mistake that technologists make, in every wave of technology, is assuming that companies in other industries are already at comparable levels of maturity to technology vendors. That these companies want to move as fast as technology vendors do into the new technology of the year and have the appetite to keep changing things. They do not. In the vast bulk of companies’ technology is viewed as a large and annoyingly necessary cost. Strategic to some (and undoubtedly strategic to the leading companies in certain segments), but not to all.

To give Anthropic their due they address some of this in the actual guide associated with the blog post. But the timelines the give are in weeks not quarters to years. Which is fine in marketing. The technical component of this problem is not the hard part.

AI is undoubtedly going to massively impact legacy systems. Personally, I left the analyst world and joined GitHub almost four years ago as it was clear something big was coming. But to quote Bill Gates “We always overestimate the change that will occur in the next two years and underestimate the change that will occur in the next ten.”

As for IBM? For as long as I have been in the technology industry commentators have been predicting the demise of IBM. Is their mainframe business under threat? Absolutely. Has it ever not been? I would be much more concerned about the revenues of the IT service providers over IBM.

• The folks at the The Cube have a great run down on the implications of Nvidia’s annoucement’s at CES in Nvidia resets the economics of AI factories, again and what it means for competitors and the broader industry. Long read, but worth your time.

• Lots of folks posted year in review pieces on LLMs and touching on coding, two I really enjoyed came from Andre Karpthy and Simon Willison.

Data Points

• JellyFish released some interesting metrics on AI Coding Tool adoption. What I like about the JellyFish data versus some of the noise out there is that it is all about enterprises, and enterprises that want to know where developers spend time. JellyFish saw GitHub Copilot and Claude Code as the first two tools in their list, and also a massive growth in AI Code Review tools in 2025.

• Interesting Gallup data on AI use in US Workplaces. The most interesting stat here is daily use, which increased by 8% to 10% in 2025. To paraphrase the “your job won’t be taken by.. “ description, “your job won’t be taken by AI, but it will be taken by someone who is using AI daily”.

Tools

• GitHub Copilot - obvousily I am going to mention GitHub a lot, but specific tools I mention during the talk include:
  • GitHub Copilot CLI
  • GitHub Copilot Coding Aggent
  • Github Copilot Code Review
• Spec Kit - SpecKit is an opensource framework released by Microsoft. It support GitHub Copilot (as you would expect), but also OpenAI Codex, Claude Code, Google Gemini and many more
• GitHub MCP Registry
• Tessl Registry

Spec Driven Development Definition

A contract for how your code should behave and becomes the source of truth your tools and AI agents use to generate, test, and validate code.

(the slide with this defintion did not load properly during the talk)

Frameworks and Data

• Dora 2024 Report (also see the 2025 report, just released last week)
• Spectrum of Approaches - the spectrum of approaches diagram in my talk was adpated from a diagram created by Richard Seroter, Chief Evangelist at Google Cloud, for a post on which Google AI developer tool in each situation. The frameing is useful for explaining tools more generically. If you don’t follow Richard you should, he is always sharing interesting pieces.
• Space Framework - research paper. DX 2025 guide.
• Is DORA, SPACE or DevEx the Right Software Metrics Framework? research piece from Gartner comparing various frameorks. This is paywalled.
• DX on Traditional Enterprise AI Adoption - lots of caveats on this figure, but its interesting.
• GitHub 20M Copilot Users

AI

• I recently set up a weekly “thought leadership”, or irreverent take on LLMs, called Making Up Shit as a Services. Joking aside, do read the About as it explains why I put this up. The initial pieces are Some of the more recent pieces include:
  • The Silent Language of Office Plants: How Foliage Drives Innovation by Google Gemini 2.5
  • Quantum Management: Leading Teams Through Parallel Corporate Realities *

• On more serious notes, we have seen a number of big waves through the AI Coding Assistants part of the industry, the biggest of these being the aquisition of Windsurf by Cognition after talks with OpenAI failed. Windsurf’s founding team and several key researchers joined Google Deepmind in an acquihire and licensing deal, which includes non-exclusive access to Windsurf’s IP for a reported $2.4Bn. Some of the very earliest employees have been very public about their dissatisfaction with the deal for non founders. TBH a lot of this deal stinks to me, but the Cognition team did the right thing by the remaining Windsurf team, and having been in a startup that went boom over a weekend thats a nice thing to see. That said I don’t have all the info, so I am drawing conclusions. Ben Thompson has a good take.

Datapoints

• GitHub (my employer), has passed 20M GitHub Copilot Users.

• RedMonk (a former employer) published their December 2024 programming language rankings, showing no change with the leading languages remaining JavaScript, Python, Java, PHP and C#. The RedMonk ranking methodology has historically depended on Stack Overflow data as one source. Stack Overflow usage continues to decline rapidly.

• Stack Overflow released their annual survey. My concerns about the demographics of this survey are well documented, but that aside there are some interesting nuggets, not least that the number of respondents has dropped by 45% - reflected in the RedMonk and Pragmatic Engineer observations above. Away from that uv, a new package manager for python, has (understandably) emerged as an admired technology. Python continues to grow in in popularity. An interesting little caveat is that Cloudflare is now more popular than AWS amoung people learning to code. I will be following this particular trend closely over the coming year.

• The Pragmatic Engineer released an annual survey, TypeScript, Python and JavaScript were the most popular languages.

• Docker released the 2025 State of Application Development survey. Lots of people use containers (unsurprisingly), although Docker do acknowledge that not everyone actually uses containers. I would argue its a defacto base choice for many new apps, and we are, thankfully, past the point where people thought they could stick an old app in a container and it was magically cloud native. Python was the most popular language among survey participants.

• Atlassian released their 2025 Developer Experience Report ,which, somewhat unsurprisingly given Atlassian sells an an Internal Developer Portal (IDP), found 98% of those surveyed planned to use an IDP. There are some non marketing facts in there as well. There is a question of scale around developer portals, and just like you should not base your plans about microservices on Netflix, you shouldn’t base your IDP needs on Spotify.

• Jellyfish, a software engineering intelligence (SEI) vendor, published survey results from analysing their user base showing AI use has increased 260% YoY with no noticeable decrease in quality. This may come as a shock to all the vendors selling tools to address all the supposed quality issues that are being introduced. Because, as we all know, all code written before AI was of the highest quality and security standards.