A Manus Rival AI Assistant "Clark" Has Arrived
An AI assistant called "Clark" was posted to Hacker News as a Show HN. Built by a solo developer inspired by Manus, it reportedly delivers a computer-use agent, browser automation, deep research capabilities, and Google tool integrations all within a single thread. Thousands of users are said to already rely on it daily.
How far an individual developer can go in implementing Manus-level agent capabilities is a hot topic in the AI agent space. We'll be keeping an eye on where Clark goes next.
Coding Agents Integrated into the Development Workflow
OneDev has announced "AI Coding Agents as Teammates." The system lets AI agents function as teammates across issues, pull requests, and CI stages. Unlike completion tools such as GitHub Copilot, the approach of embedding agents directly into the development workflow itself is gaining traction.
Along similar lines, an AI assistant for ArgoCD has also been released on GitHub, accelerating AI integration in the DevOps space.
The Hidden Cost of AI-Assisted Development: Cognitive Fatigue
An article titled "The Cost of AI-Assisted Development: Cognitive Fatigue" has sparked discussion. While AI tools do speed up development, it points out that the burden of reviewing and verifying output grows — and developers run into unexpected "walls." At a time when productivity gains get all the attention, this is an important perspective that puts the spotlight on the fatigue cost borne by developers.
llama.cpp: A Critical Fix for Agent Workflows
llama.cpp build b9978 fixed a context checkpoint bug that affected agent workloads. Previously, a new checkpoint was created on every agent turn, and rewinding the context during tool-call loops wiped out all checkpoints. With this fix, long-running agent sessions are expected to get significantly faster.
This is a big win for anyone building agent workflows with local LLMs.
Running Qwen3.5-122B on a Mac Studio: Three Bug Fixes That Optimize Long-Context Inference
The LocalLLaMA community on Reddit shared a case study of running Qwen3.5-122B on an M3 Ultra Mac Studio (96GB) and resolving the slowdowns that occur during long-context agent coding. By fixing three bugs — prompt instability, the streaming interrupt path, and checkpoint contamination — they reportedly cut prefill time from minutes down to sub-second.
The fixes have been open-sourced as a qMLX fork, making them a useful reference for developers working on hybrid attention caching optimizations.
News from Japan
Hitachi on the Role of "Humans" in the AI Era**: Hitachi, which is pushing AI adoption as it modernizes legacy systems, emphasizes that "human capabilities will matter more than ever." As AI adoption advances, the company is reasserting the importance of human-led quality assurance and judgment.
Scene Workspace Automates Assembly Process Creation from 3D CAD with AI**: A feature co-developed by Scene and DENSO's Kiki division that automatically generates assembly processes and work instructions from 3D CAD data. An initiative to connect design through manufacturing on a single platform.
freee MCP Now Usable from Claude Code**: An MCP server that lets you operate accounting and HR APIs in natural language was introduced on Qiita. With instructions like "Show me this month's sales," you can directly work with freee data.