Anant Jain

AI Coding Tools (May 2025)

Tech

Over the past couple of months, I took a short break to explore the rapid developments in AI-powered coding. Software engineering is undergoing a significant shift, and I wanted enough hands-on experience to understand where the tools were genuinely useful—and where the excitement was running ahead of them.

My goal was simple: try as many AI-powered development environments, coding agents, and similar tools as I could. I wanted to understand the innovation frontier and, if I got lucky, find both an opportunity and a potential co-founder for a startup in this space.

This post records what I learned and the opportunities I think remain. As of May 2025, the tools I surveyed fall into three broad categories based on how they fit into a development workflow. The boundaries are imperfect, but the categories make their different interaction models easier to compare.

1. AI-assisted development environments

These tools boost individual developer productivity through a local development environment. They function as AI pair-programming partners:

  1. Synchronous and iterative
  2. Best for higher ambiguity/complexity tasks
  3. GTM: Individual adoption, leading to a bottom-up go-to-market motion
  4. UX: An app installed on your computer

I would further divide this category into three smaller sub-categories:

Terminal-based:

  1. Warp: This isn't just a tool within your favorite terminal (typically, iTerm2), but a whole new AI-powered terminal that comes with a coding agent. I didn't use Warp's coding agent extensively, but it successfully replaced iTerm2 for me—the ability to describe what you want to do in plain English without memorizing Unix commands comes in handy once in a while.
  2. Claude Code: This is my favorite tool on this list. It's the best-designed terminal tool I've seen. The Claude Code agent works well, and I like how it uses Claude.md as a shared scratchpad in the repo to remember codebase preferences over time. My only wish is that Anthropic would open-source it, especially after the code leaked via source maps and someone created a multi-provider version that Anthropic recently DMCA’ed as well 🤷🏻‍♂️
  3. OpenAI Codex: Claude Code's first-mover advantage was significant for me, and Codex didn't feel like an improvement in my limited experimentation. Massive props to OpenAI for making it open source, though!
  4. Aider: While I didn't explore this as much as intended, it has received considerable praise in open-source-loving Reddit communities.

Editors (IDEs):

  1. Cursor: My most-used tool on this list. All things considered, I believe this is the strongest product and team in the IDE space. I just wish their growing pains were not so evident.
  2. Windsurf: My initial experience a few months ago was underwhelming, but trying it again this past month was decent. While it wasn't better than Cursor for my use cases, they claim superior performance with larger codebases.
  3. Zed: The latest entrant in this category, created by the Atom editor team. While incredibly fast (built from scratch in Rust), betting on speed as a differentiator seems risky when the improvement margin over VSCode-based tools isn't substantial.

VSCode Plugins:

  1. RooCode (forked from Cline): I quite enjoyed RooCode, and it is my favorite in this category. I especially liked having the Architect mode separate from Code mode. However, the "bring your own API key" approach is expensive—especially when tackling ambiguous problems where growing context diminishes model performance while token usage and costs keep climbing.
  2. GitHub Copilot: Though it has lost momentum since 2022, it remains well-positioned to capture the enterprise market through Microsoft's proven bundling strategy.
  3. AugmentCode: Deserves a mention for being on top of the SWE-Bench Verified leaderboard as of today.

2. AI “Engineers”

AI agents that handle complete tasks independently. These tools operate asynchronously with a "ticket to pull request" workflow, similar to delegating work to a junior engineer:

  1. Asynchronous processing with single or limited interactions
  2. Ideal for straightforward, well-defined tasks
  3. GTM: Team-wide adoption, implying an enterprise (top-down) sales motion
  4. UX: Ambient agents with Agent Inbox (Dashboard)

Here are the key products I experimented with in this category:

  1. Devin: This team is extremely impressive and on a great trajectory. While the initial version didn't amaze me, v2 was significantly better—provided you set appropriate expectations.
  2. Charlie: I got a chance to try out the private beta, and have to say, Charlie is truly impressive. The team pivoted into this space earlier this year, and from my limited testing, they seem to outperform Devin.
  3. Factory: I haven't tested this one. In such a competitive space, I'm generally skeptical of products that require talking to a sales AE before trying them out.
  4. [Update 05/16] OpenAI's Codex: Powered by codex-1, a version of OpenAI o3 optimized for software engineering, it can perform tasks in parallel, such as refactoring, bug fixing, and documentation. I have yet to test it extensively.

3. Prototyping tools

I include this category here, but it wasn't the focus of my exploration. I've used v0.dev extensively, but only to work with PMs and designers as we explored ideas—as a Figma replacement. The three other notable tools here are Bolt.new, Lovable, and Replit. There are minor differentiators between the four, but each is trying to provide hosting or sell a Figma replacement to enterprises to make its revenue stickier.

What’s missing?

Here are a few products/ideas that I’m excited about in this space:

  1. Building your own agents: I think we’re very close to nailing down the UX of building your own agent for non-developers. Folks broadly realize that this could be useful (e.g., YC’s Internal Agent Builder RFS for Summer ‘25), and I wouldn’t be surprised if one of the current AI agent frameworks pivots into this.
  2. Staff engineer agent: I think we’re missing an agent fine-tuned to take larger-scope, ambiguous problems and turn them into tractable plans. In particular, the UX should resemble what we humans do when we kick off significant projects: write a doc → share it with the team and iterate through comments → get approval from the decision maker → turn it into a plan with milestones and issues. Devin is well-positioned to incorporate this, but it seems focused for the most part on building a “junior engineer that codes.”
  3. Issue tracking: Linear remains my absolute favorite here—it's great to see the team setting up an MCP server and maintaining clean APIs for integrations. I think Linear can become the dashboard and orchestration layer for AI agents. More broadly, if AI-native products disrupt traditional SaaS applications, project and issue tracking looks like a particularly ripe space.

The common thread across these ideas is that implementation is getting cheaper while context, coordination, and judgment remain scarce. The most interesting tools may not be the ones that write the most code, but the ones that help people and agents decide what to build, divide the work, and keep the result coherent.