Transcription
A brand new coding agent just outscored GPT5 codecs on Terminal Bench. It also beat Claude Code on SWEBench, hitting 74%, the highest number we've seen yet. And it's not from OpenAI or Anthropic, but from Abacus AI. They call it Deep Agent Desktop. And it might just be the smartest coding agent released so far. So, let's talk about it.
All right. So the clearest way to see it is in the benchmarks on terminal bench. Deep agent desktop scored 48.75%. For context, GPT5 Codeex came in at 42.8%. Claude Code Opus sat around 43.2%. Claude Code Sonnet 4 hit 35.5% and Goose landed at 45.3%. So, Deep Agent is clearly ahead there. Not by a tiny margin either.
Then on SWEBench verified, which is basically the standard for automated bug fixing and real-world software engineering problems, Deep Agent scored 74%. Again, GPT5 Codeex hit 72.8%. Claude Code Sonnet 4 was 72.7%. Opus was 72.5% and the older Claude Code Sonnet 3.6 7 dropped all the way down to 62.3%. And that means in both head-to-head comparisons, this new system is at the top. And these numbers matter because SWEBench isn't about toy tasks. It's about fixing real GitHub issues pulled from real projects.
So what exactly is Deep Agent Desktop? Well, it's not just a single model. It's actually a full desktop suite that combines three different modes. A CLI agent, a code editor agent, and a chat mode that can access different models, including Claude, Gemini, and GPT5. And here's the clever part. It also comes with its own testing agent. That's right. It doesn't just write code, it tests the code itself to make sure the solution works. That's something most other systems don't offer out of the box, and it's probably a big reason why the benchmark scores came out so strong.
Let's start with the CLI mode because that's where they've been showing off the most. They literally branded it as the fastest way to code. The workflow is simple. You type your request in the terminal and Deep Agent builds it out in real time. In one of the demos, someone asked it to build a quick snake web game with a retro Nintendo vibe, complete with gamified badges and awards for different levels and scores. The output wasn't just some basic grid with a moving dot. It was styled like a 90s console game. Visually appealing, interactive, and smooth. The fact that an agent can spin that up with just one prompt in the terminal is pretty wild.
Then they went further. Another prompt asked for a working LinkedIn clone. They even gave it a name, ConnectHub. The requirements weren't simple. Users had to be able to sign up, log in, connect with each other, post content, and all of that had to run on a modern tech stack with Django on the back end, a proper database and a polished front end. Deep Agent CLI handled that, building out a full web app structure. Again, that's not the kind of project you expect from a one-off prompt, but it shows the agent's ability to handle real-world requests. And in another demo, they connected it directly to a GitHub repo called the writer's corner. This is a community site for novel writing. And the task was to add a leaderboard based on engagement, not just likes, but also factoring in recency and the number of comments. So the ranking feels dynamic. Deep Agent went through the repo, extended the app, and implemented the leaderboard logic. That's not something you can fake in a benchmark slide. It's real repo manipulation, and they actually showed it working.
Now, if we move over to the code editor mode, this is more like a full IDE experience powered by AI. The demo here was also impressive. They uploaded a resume as an image and asked Deep Agent to create a personal website from it. The system read the resume, pulled out the details, and then built a modern, polished website showcasing the candidate's information. That's a huge upgrade from just generating an HTML template because it had to do OCR on the resume, extract structured data and then design around it.
Another example in the code editor was even more ambitious. They asked for an advanced holistic guide for vibe coders structured into beginner, intermediate, and advanced levels. It had to cover principles, frameworks, and rules of thumb for scalable, maintainable, and robust applications. The guide also had to include mental models, learning paths, and practical examples showing not just how to write code, but how to work with AI generated code, audit it, and improve it. The agent didn't just dump text. It built this as a modern interactive website with smooth animations, a clean design, and even a curated list of high-quality blogs and resources at the end. That's a full product experience, not just a wall of markdown.
Then there's chat mode, which basically turns Deep Agent Desktop into an all-in-one AI assistant that lives on your desktop. This mode lets you talk not just to Deep Agent's internal system but also to external models like Claude, Gemini, and GPT5. The point here is flexibility. You don't have to switch between apps. Everything is under one roof. The prompts they showed in chat mode were wide-ranging. One person asked for a 90-day playbook on becoming a LinkedIn and Twitter influencer in software development. The agent produced a detailed plan with positioning, audience targeting, content pillars, repurposing workflows, posting cadence, community engagement strategies, growth experiments, tools, and KPIs. It even laid out a calendar with daily and weekly tasks, measurable targets, sample posts, and checklists. That's way beyond what most people would expect from a casual "help me grow online" request.
Another prompt was about running open-source indie game repos from GitHub. Deep Agent gave step-by-step instructions for setting them up locally. Then there was a space-related one. Someone asked how spacecraft land on Mars so accurately. The agent broke it down in a cinematic style, walking through trajectory calculations, mid-course corrections, and the tech that scientists use to guide landers with pinpoint precision. Again, it wasn't just surface-level explanations. It had the detail you'd expect if you were trying to actually remember the process.
What's important here is that Deep Agent Desktop isn't positioning itself as just another model. It's a full package and the company is smart about making it accessible. They priced the basic tier at $10 per month, which is way lower than what most developers spend on coding assistance or premium IDE plugins. They even mentioned they'll be hosting weekly competitions where people can showcase their human-AI collaborations and win $2,500. That's a clever growth play because it incentivizes the community to actually use the product in creative ways and share it publicly.
Now, benchmarks aside, the framing from Abacus AI is pretty clear. They want Deep Agent Desktop to be the go-to coding agent. The message in their launch was direct: "the smartest coding agent yet," and the data does back that up, at least for now. GPT5 Codeex and Claude Code have been the heavyweights for months, and this is the first time we've seen another system outperform both at once.
Now, if you're thinking about trying it out, they've kept it really simple. You just run one command in your terminal: `npx deep-agent cli`, and the agent is ready to go. From there, you can hop into the code editor if you want the full desktop IDE setup, or stick with chat mode to switch between models like Claude, Gemini, and GPT5. The whole idea is that it's all in one place so you're not juggling different apps or subscriptions.
So, yeah, Deep Agent Desktop just officially arrived and right now it's sitting on top of the coding benchmarks. If you're a developer or even just someone curious about where AI is going with software engineering, this is definitely something worth paying attention to.
And that's it for this one. Don't forget to drop a comment, hit that like button, subscribe if you haven't already, and I'll catch you in the next video.