Follow ZDNET: Add us as a favorite source On Google.
ZDNET Highlights
- Moonshot AI pushes autonomous coding to new limits.
- Designs and builds full-stack apps from AI signals.
- Persistent agents run around for days handling real tasks.
Yesterday, Moonshot AI announced Km K2.6The latest version of its open-source AI model. This release has enhanced coding capabilities, longer multi-step operation execution, and agent swarm capabilities (which doesn’t sound scary at all).
Also: The best free AI for coding – only 3 available now
Based on a reinterpretation of the OpenClaw AI assistant approach to automated AI processing for complex, real-world workflows, the company is doubling down on a “seamless AI coworking experience.”
Long-horizon coding performance improvements
At the core of the Kimi K2.6 release is a substantial improvement in long-horizon coding performance. Long-horizon coding is another way of saying that AI can perform a very long series of steps without human supervision.
Think of the difference between short-horizon and long-horizon as an employee who you have to check in on every 15 minutes, and an employee who you can just give an assignment to and know that what you need will be on your desk tomorrow morning without any fuss or hassle.
Also: 7 AI coding techniques I use to ship real, reliable products faster
Moonshot uses a SysY compiler project as an example of a long-horizon assignment. SysY is a minimal C-like language used to teach compiler design to students. Kimi K2.6 designed and built a complete SysY compiler from scratch in 10 hours, passing 140 functional tests without human input. It said that this work is equivalent to the work of four engineers for two months.
Undoubtedly, this is a remarkable achievement. But Moonshot isn’t alone in using AI to create compilers. Anthropic reported this in February It created a complete C compiler (Not just the cut-down Training Wheels version) Using your Opus 4.6 model.
The Anthropic project performed quite well, but when the agents completed the complex task of compiling the Linux kernel, it hit a bottleneck, causing them to run into the same bugs, overwrite each other’s work, and break existing functionality by adding new features.
I’m guessing that the choice of SysY on the part of the Kimi developers was to keep the overall complexity low, and this new model will likely hit the same set of constraints that Anthropic faced.
Moonshot says the K2.6 model exhibits strong generalization (meaning it is able to handle new and unexpected situations) in languages including Rust, Go, and Python. It also reports that the new model demonstrates reliability in front-end, DevOps, and performance optimization tasks.
Expanding from coding to design and construction
Coding output isn’t the K2.6’s only big trick. The model is capable of performing user interface design tasks and then generating coding output from that design. It enables non-coders to create complete web applications from prompts, including look and feel. It provides assistance to developers who may not have design expertise.
Also: I Tried to Save $1,200 by Vibe Coding for Free – and I Immediately Regretted It
Going back to the long-term claim discussed earlier, Moonshot demonstrated full-scale project potential by creating a series of websites. The company reported that Kmi K2.6, “identified 30 restaurants in Los Angeles without official websites, then automatically created high-converting landing pages for each. These pages included booking functionality, with all information seamlessly synchronized with their database.”
Swarms of agents, active agents and persistent execution
According to Zilin Yang, founder of Moonshot AI, “By orchestrating 100 or even 1,000 sub-agents in parallel, we can complete complex tasks within time frames tolerable for the real world.” It calls it “agent swarming”.
I don’t know. I’ve probably watched The Terminator too many times, but while I can see the practical benefits, the idea of a bunch of AI agents is just weird.
The company reports, “It seamlessly coordinates heterogeneous agents to combine complementary skills and broad search capabilities with in-depth research, as well as large-scale document analysis paired with long-form writing, and multi-format content creation executed in parallel.”
It adds that, “This compositional intelligence enables Swarm to deliver end-to-end output spanning documents, websites, slides, and spreadsheets within a single autonomous run.”
The KMI K2.6 model now supports autonomous agents working continuously across applications and workflows. This release also improves API interpretation, long-term stability, and security awareness.
The company demonstrated a K2.6-supported agent that “operated autonomously for 5 days, managing monitoring, incident response, and system operations, performing persistent context, multi-threaded task management, and full-cycle execution from alert to resolution.”
Too: MIT study shows AI agents are fast, loose, and out of control
Another capability added in Kimi K2.6 is what the company calls “Claw Groups”, which enable multiple OpenClaw-style agents running on different devices to collaborate with a shared context. There is a central coordinator that dynamically assigns tasks and resolves failures.
Moonshot AI says that all of this becomes a form of collective intelligence. It says, “We are moving beyond simply asking AI a question or delegating a task to AI, and entering a phase where humans and AI collaborate as true partners – combining strengths to collectively solve problems.”
As long as the agents don’t go and invent time travel, we’re probably safe. For now.
Would you feel comfortable letting an AI agent work continuously for several days, managing the system on your behalf? Let us know in the comments below.
You can follow my daily project updates on social media. Be sure to subscribe My weekly update newsletterAnd follow me on Twitter/X @davidgewirtzon facebook Facebook.com/DavidGewirtzon instagram Instagram.com/DavidGewirtzon bluesky @DavidGewirtz.comand on youtube YouTube.com/DavidGewirtzTV.
