1Password Refactors Its Monolith with AI Agents — Lessons from a Real Production System
1Password, the password management company, has shared its experience using AI agents to refactor its own product's monolith. Migrating from a monolith to microservices is a challenge many companies face, and the lessons drawn from this real project—how much can be delegated to AI agents and where humans need to step in—offer useful reference for development teams grappling with the same issue. According to the report, AI agents excel at understanding code and performing mechanical rewrites, but human review remained essential for architecture-level decisions and boundary design.
Toyota Finance Shortens Inquiry Handling to 4 Minutes with AI Agents
Toyota Finance has introduced AI agents into its customer inquiry operations, cutting per-case handling time from 13 minutes to 4 minutes. What makes this notable is the focus on automating "non-routine work" rather than simple FAQ auto-response. When selecting an AI agent, the company reportedly compared multiple services and prioritized whether the tool could be integrated into its actual business workflows. As a case study of AI agent adoption within a major Japanese corporate group, the report covers everything from the implementation process to effectiveness measurement in detail.
Qwen 3.6 35B Reaches Practical Speeds on Consumer GPUs
In the local LLM community, a report of running Qwen 3.6 35B A3B on two RTX 5060Ti cards (32GB VRAM total) and achieving 90 tokens/s has been drawing attention. Using LM Studio with Q4 quantization, the model runs with full context. Around the same time, reports also surfaced of cluster builds with two RTX 3090s, signaling that we have entered an era where 7B–35B parameter-class models run at practical speeds on consumer GPUs. Meanwhile, on the Hugging Face trends, uncensored fine-tuned models based on Qwen3.6-27B are also ranking near the top, showing that the community remains as active as ever.
"The Other Half of AI Safety" — The Overlooked Risks
An article titled "The Other Half of AI Safety" that gained traction on Hacker News argues that AI safety discussions are skewed toward a subset of risks. Mainstream AI safety debates tend to focus on superintelligence and existential risks, but the article contends that the "other half"—risks already affecting people today, such as misinformation, privacy violations, labor market disruption, and widening social inequality—is being neglected.
Separately, Zvi Mowshowitz has also sounded the alarm about the current state of cybersecurity and AI governance. His point that AI systems are being deployed on insufficient security foundations while governance frameworks fail to keep pace with technological progress is something companies should keep in mind as they push forward with AI adoption.
A Quantitative Comparison of the Big Three AI Coding Tools: Claude Code vs Cursor Agent vs Codex CLI
An article published on Qiita ran the same refactoring task through Claude Code, Cursor Agent, and Codex CLI, quantitatively comparing speed, accuracy, cost, and the number of operation steps. The conclusion: "there is no universal tool—the right answer is to use each according to the task's characteristics." The finding that each has distinct strengths and weaknesses, with certain tools suiting large-scale changes and others better suited to fine-grained fixes, offers practical guidance for engineers using AI tools in their day-to-day development work.
"Cognitive Surrender" in AI Development and Maintaining Debugging Skills
On Zenn, a discussion of "Cognitive Surrender" in AI coding has been gaining attention. The term describes the state in which, as you have AI write your code, its output quietly substitutes for your own. The debate spread from a blog post by Addy Osmani. In particular, how to keep your debugging skills intact while coding with AI is an urgent challenge for engineers.
There is also an argument that vague rules like "be careful" are meaningless. Writing "be careful" in rule documents such as CLAUDE.md won't make it happen unless it is physically enforced—the view that an approach which makes violations "physically impossible" through linters and tests is more effective also resonates with designing trust boundaries for AI agents.