“Agentic” coding tools are the new hot AI-wrapper product. They seem to promise that they will make your developers super-humanly productive by turning them into managers delegating and approving the work of as many AI coding assistants as you can afford.
They are also spoken of as the next step in the evolution of programming. As we went from filling memory by flipping switches to manually punching cards to encode machine instructions then on to assembly language and from there to structured programming languages that required a compiler to generate machine code, now we will all be programming in English (or your preferred human language) via conversation with AI.
This new conversational style of “programming” is also causing people to predict the end of the IDE as the new agentic coding tools do away with or simplify text editors as part of their feature set.
Cursor was the editor with built-in AI. Now you have tools like Factory and Jules that reduce the editor to a minimal box where you can make basic changes if you really must. If you have a problem with your agentic AI assistant’s code, or if you just want to explore what they’ve written, you’ll need to tab away to your old IDE.
AI-assisted coding is the second killer app after ChatGPT (which took just 2 months to reach 100 million users) and model providers are leaning in hard to capture this market, shifting the training of their models to emphasise coding ability and coding processes.
And on the product side, the industry saw the valuations of Cursor and the purchase price of WindSurf and started pumping out their own variations and visions for the future of AI-assisted coding.
Below we run you through the main contenders for agentic coding assistants. “Agentic” is slowly gravitating towards meaning multi-agent systems performing multiple (if not hundreds) of actions on their own to complete a task. Coding assistants are all heading that way. Having many agents doing focused tasks and expending higher levels of compute is a clear strategy for getting better results out of AI.
But these coding tools are mainly single agent assistants and “agentic” here means that the coding assistant will decide what to do itself across many, even hundreds, of actions.
Some developers simply run multiple instances of these single agent assistants simultaneously. Here is the CPO of Anthropic, makers of Claude, explaining that this is exactly what happens in Anthropic, where developers have become “orchestrators” of Claude, and yes, that is going to impact hiring practices.
The Agentic Coding Tools
Claude Code
The first widely used agentic coding assistant, Claude Code was released in 2025. Anyone who has been closely following the tech for more than a year will recognise the influence of the open source AI coding tool aider. Claude Code took the terminal-based, conversational model and added MCP-based tool calling, giving the assistant more actions to perform, including interacting with files, searching the web for solutions, pushing changes to your git repository and anything else you wanted to wire it up to.
If you wanted to look at the code you had to switch to your IDE. For VS Code and the like you could choose to run Claude in a terminal window and watch it work while giving it directions.
Running multiple instances of Claude Code in a terminal using a session manager like tmux became a power move for developers who could afford the expense of all the tokens. This practice was codified in tools like Claude Squad.
Devin
Devin made a splash when it was announced in March 2024. Its big selling point was that it was built by a team of competitive coders, who obviously must know a thing or two about software development. Unlike Claude Code, which anyone who could be bothered to sign up for an Anthropic API key could access on a PAYG basis, it was infamous for being expensive and hard to get access to. It became generally available in December 2024.
With the release of Claude Code in the following February, which gave developers a new sense of just how expensive coding can be when every action can consume 100K+ tokens, Devin no longer seemed over-priced.
Devin has an in-house fine-tuned model for code generation. It also uses dedicated agents for different purposes (editing files, interacting with the command line, etc) that can interact with each other to get things done.
In May 2025 OpenAI announced Codex, their own dedicated AI coding assistant running on their models. Codex is cloud-based and can work with your repositories. It is only available in ChatGPT Pro and ChatGPT for Teams.
At the same time OpenAI also announced Codex CLI, an open source Claude Code clone that the community quickly updated to make it work with other model providers and inference services.
Jules
Google announced Jules, their cloud-based coding assistant at Google I/O in May 2025. It is powered by their SOTA Gemini models.
Jules can connect to your repositories and it uses a notion of “tasks” to allow you to direct it to work on several things at once. It is still in early beta and provides 60 actions per day for you to try it out.
AmpCode
AmpCode looked at how developers were using Claude Code, especially developers running multiple instances of Claude Code to do more at once, and built an interface around that idea. They extended it, calling multiple instances “Threads” and making it team based, so everyone involved can see what is being worked on. They recently let the agents in a Thread spawn sub-agents that can run in parallel.
AmpCode is available as a VS Code plugin and as a node-based CLI tool.
Factory
Factory is the latest AI coding tool to come out of stealth mode. It is a browser-based tool, like Jules, but unlike it Jules it also has a “bridge” app that runs on your machine, allowing it to access local files and the command line.
Factory uses the idea of “Droids”, which are each a specialised collection of prompts and tools. There are Knowledge, Product, Code and Reliability Droids.
The idea with Factory is that you have multiple sessions running, each in its own browser tab or window, each using a particular type of Droid to perform tasks.
With the right tool permissions, Droids can update your local or Github repositories directly. And the interface lets you work as a designer and code reviewer instead of a programmer. You will want to pop out to your IDE opened in your repository if you want to explore any changes or make your own fixes.
Different Interfaces But Same Models
Each of the tools we’ve covered has their own take on how AI-based coding is going to be performed. Some are more different than others, but we are in the early, exploratory stage of this paradigm.
In trialing these tools for this article (except for Devin) one thing was obvious: no tool is magical. They all feel the same and that is because they are all built on top of the same 3 model providers.
The SOTA models are all pretty close together in the evals. No bespoke interface, despite any lengthy, heavily tweaked prompts, is going to extract better code out of the models than any other product.
Pick Your Coding Assistant
The only way to get better code out of a model is more compute – more tokens, more context, more runs at the same task. It is pay-to-win, but with open source tools like Cline and less expensive, high ability coding models like DeepSeek, you do have cheaper options.
An effective coding assistant does more than just generate code, it also performs actions. There are differences in models’ abilities to choose, execute and respond to actions, with Claude being the leader in this regard, but it is a feature all model providers are training for so you can expect the gaps to close in the near future.
With models matching on quality, and tools racing for feature parity, and everyone competing for your $$$, it’s a good time to be trialling tools and seeing what works best for your team.
Your AI coding strategies can be transplanted from one tool to another. Their output is raw code in your repositories, so there is no lock-in.
Coding is going to be AI-assisted. There is no avoiding it. Start working with the tools now so your team can evolve alongside the technology and not be left behind.
