The dawn of composition
Claude 3.7 ushers in the era of workflow in software development
The AI roller coaster continues unabated. Last Monday Grok 3 was released. Yesterday we got Claude 3.7. It’s not inconceivable we get GPT-5 next week.
I’ve been a fan of Claude for many months. And Claude 3.7 does not disappoint. It still feels like Claude. It has the same concise, pleasant, focussed, not-a-show-off character that I loved in Claude 3.5. But the base model is smarter.
For me this graph tells the story. Anthropic have given Claude basic tools to play Pokémon. Look at how poorly Claude 3.0 Sonnet does compared to 3.7. And then reflect that Claude 3.0 Sonnet is from March 2024. Just eleven months ago.
As Anthropic note:
As you can see, the previous versions became stuck very early in the game, with Claude 3.0 Sonnet failing to even leave the house in Pallet Town where the story begins.
But Claude 3.7 Sonnet’s improved agentic capabilities helped it advance much further, successfully battling three Pokémon Gym Leaders (the game’s bosses) and winning their Badges. Claude 3.7 Sonnet is super effective at trying multiple strategies and questioning previous assumptions, which allow it to improve its own capabilities as it progresses.
Claude the software dev
And the software benchmark scores are better too. A promising sign for a model that is generally accepted as being the best software development model (I have a slight preference for o3-mini as it seems better at one-shotting Rust code, but I’m excited to learn what Claude can do).
Anthropic is touting Claude 3.7 as the first hybrid reasoning model. It can behave like a traditional model (just like all previous versions of Claude). But, like o1, o3 and DeepSeek, it can also spend more time thinking. And, it shares its chain of thought (though interestingly, not in Claude's authentic voice...)
So what can it do? It seems increasingly obvious that we need to lean into using AI as both teacher and assistant. So I got it to create an introductory Python course…
I left it generating and went away and did some other things. And when I came back I did a double-take. I got this artefact (which you can try here).
Now, in fairness, Claude got tied in knots on its first attempt. But the second attempt is remarkable. Claude 3.7 is amazing; suddenly it’s clear why xAI rushed out Grok 3 last week - it was their only chance for a week in the spotlight before being overshadowed by Claude.
Claude Code
And Claude comes with other goodies. Through the developer API it’s possible to control how much time Claude spends on thinking - something we’ve never had control of before. And then there’s the jewel - Claude Code. It’s worth watching the video here. It’s a complete LLM based development tool - codebase analysis, feature enhancement (including writing tests), iterating to fix compile errors and, GitHub integration. Wow.
This is the start of tools appearing to support the workflow development process I outlined last September. A reimagined software development flow - where software development smarts are added to the LLM - rather than retrofitting AI to existing tools. Cursor became the fastest-growing SaaS company to scale from $1 million to $100 million in annual recurring revenue within 12 months. But they may soon find themselves irrelevant as the software development landscape continues to rapidly evolve. And I doubt Anthropic will limit themselves to just code. How long before we get Claude Word? Claude Presentation?
The three eras of AI-assisted development
Over the past few years a pattern has emerged in the development of AI-assisted development:
Era 1: The Assistants (2022-2023) This was the GitHub Copilot era - AI that could help complete lines of code or suggest simple functions. These tools were intra-method focused, good at predicting what you might type next, but ultimately still requiring developers to think about architecture, testing, and integration. They saved time but didn't fundamentally change how we built software.
Era 2: The Composers (2023-2024) This is where we've been with models like Claude 3.5 and o3-mini - tools capable of producing entire modules (500-1000 lines) of coherent, working code. These models understand patterns, libraries, and frameworks. They can generate entire components, but still need human guidance for system-level integration. This era has already dramatically increased productivity for those willing to adapt.
Era 3: The Workflows (2024-?) Claude 3.7 and Claude Code are the first true glimpses of this era. Here, AI doesn't just write code - it manages the entire development lifecycle from requirements through deployment. The AI can reason about codebases as a whole, detect bugs across multiple files, write tests that cover edge cases, and refactor with system-wide awareness. This era is where the traditional role of the software developer transforms completely.
We're just entering Era 3 and the implications are still unclear. The transition from Era 1 to Era 3 has happened in barely 24 months - far faster than most of the industry has been able to adapt. This suggests that by 2026, we may well be in an entirely new paradigm that today's tools hardly hint at.
And so?
The significance of Claude 3.7's release isn't its improved reasoning or even its coding capabilities - it's what this reveals about the pace of change. Look at how things have advanced over the past few years. And then consider where they are headed. My o3-mini experience 10x’d my productivity. What will Claude Code do? And where will we be in another 12 months time?
Claude 3.7 makes it easier than ever to develop software. More and more people can now build code. We’re at the start of an explosion of creativity and innovation. By the end of 2025, there will be more software written by AI than by humans. And companies that have failed to transform their development workflows to centre around AI will be at an insurmountable competitive disadvantage - unable to match the speed, quality, or cost-efficiency of their AI-native competitors.
The question isn't whether your role as a developer will change - it's whether you'll be orchestrating a team of AI agents to build systems of unprecedented complexity, or watching from the sidelines as others do. The coding skills that defined careers for decades are being commoditized at breathtaking speed, but the opportunity to create has never been greater.
What will you build when code is no longer the limiting factor?