AI Coding Tools: The Ultimate Guide in 2024

92% of developers are now using AI coding tools, and it is obvious why: it improves their productivity. In some cases, it even makes them twice as productive:

The effectiveness of these tools has led to a gold rush in AI coding tools, with many even declaring the imminent demise of human software engineering. 

In this article, we’ll dive into the current frontier of AI coding tools, discuss what’s on the horizon, and finish with some thoughts on how you can take advantage of this trend. 

The Current Market

Most AI coding tools are built on top of GPT-4, and there still isn’t a GPT-5. So it’d be fair to assume that the cutting edge is still GitHub’s Copilot and Cursor. But, if you look a bit deeper, you’ll see that there has been a good amount of innovation even without meaningful improvements in the underlying model. 

Take Codium, for example. They have their own pair programmer, but unlike Copilot, they also have an AI agent that speeds up pull request review and a cover agent that automates test coverage. 

Or take Supermaven and their two main innovations. The first is their 300,000 token context window, which is by far the largest option on the market (for comparison, Copilot’s is 8,192). This means that it can better understand your codebase, and thus produce better results. The second is that because of their custom model architecture (they don’t use a transformer), they are able to generate code 3x faster than Copilot and 7.5x faster than Cursor. The result is a tool from a small startup that is legitimately competing with Unicorns.

In the open-source world, you have tools like Code2Prompt. One of the weaknesses of regular chatbots is that they don’t integrate with your codebase. So, Code2Prompt converts your entire codebase into a single prompt that LLMs can easily understand. You can then upload this prompt to a chatbot like ChatGPT or Claude and work with it there. Basically, it turns chatbots like ChatGPT into copilots.

Niching down even further, there are now a bunch of tools that do one specific thing really well: 

• CodeWP is a copilot built explicitly for WordPress

• AI2sql crafts efficient and accurate SQL (and no SQL) queries. 

• Prisma Assistant lets you interact with your prisma schema conversationally. 

• ReactAgent generates React components from user stories and existing components. 

AI coding tools are already very useful and continue to improve even without significant improvements to the underlying model. However, these improvements pale in comparison to the potential of AI coding agents, which is where much of the financial and intellectual power is now going.

The Frontier

The AI agent rush began in March with the release of Devin, “the first AI software engineer”, by Cognition Labs. What Cognition did is imbue an AI coding tool with reasoning, enabling it to complete tasks that require actual decision-making. In other words, Devin is an AI coding tool with at least somewhat of a brain, as it can solve 13.86% of real-world engineering issues.

The success of Devin has led to an explosion in AI coding agents, but probably not in the way you’d expect. With the exception of Amazon’s Q Developer, an AI agent that can solve real-world issues 13.82% of the time, the cutting-edge is dominated by open-source projects:

• OpenDevin and Devika aim to match and eventually surpass Devin’s abilities. 

• CodeR uses a multi-agent framework and pre-defined task graph to solve 28.33% of SWE-Bench Lite (problems that are a bit easier) issues. 

• SWE-Agent uses simple LM-centric commands and specially built input and output formats to make it easier for the LM to browse the repository, view, edit, and execute code files. This allows the LM to solve 12.47% of real-world issues. 

• AutoCodeRover’s analysis and debugging capabilities enable it to resolve 15.95% of real-world issues. 

The reason for this open-source domination is simple: AI agents aren’t good enough yet to be productized. It’s hard to sell a product that only works 15% of the time. So, while they are undoubtedly impressive, there is still a long way to go until AI agents are able to replace your software engineers.

The Business Opportunities

With the AI coding tool market projected to grow to $12.6B by 2028, this field is definitely worth looking at. However, if you want to compete with Copilot, which already has over 1.3M users and Microsoft’s financial and technical backing, you’ll have to think outside of the box.

The most obvious avenue would be to build an AI coding agent. This is a multi-billion dollar opportunity if you can do it well, as evidenced by Cognition Lab’s $2B valuation after just 6 months in business. However, because it’s on the cutting edge, you’re going to need an absolutely cracked team. The founders of Cognition are some of the best developers in the world and legends in the competitive math scene. You won’t have any hope of competing without a comparable team. 

So, for the rest of us mortals, the opportunity is in niching down. This might look like building a domain-specific tool like AI2sql, building a tool that integrates with a no-code framework like CodeWP, or building a tool that “knows” more obscure languages, a well-known weakness in the most popular AI coding tools. 

The key here is the same as countless other indie projects: finding and serving an overlooked market. If you can do that, you’ll be able to build a profitable product without being on the absolute frontier or relying on VC funding.