Here is a roundup of AI news that drew attention from international media and communities on August 2, 2026 — covering model behavior, tools for orchestrating agents, reproducing voice AI locally, controlling agent payments, and practical know-how shared in Japanese.
DeepSeek-V4-Flash-0731: Reasoning "Low" Beats "High" in a Reversal
DeepSeek's lightweight model "DeepSeek-V4-Flash-0731" lets you choose reasoning intensity from four levels (None, Low, High, and Max). But when a developer tested each mode 20 times, they confirmed a counterintuitive result: "Low" consumed more tokens and reasoned longer than "High".
Low**: about 1,227 tokens on average (about 874 of them reasoning)
High**: about 606 tokens on average (about 411 of them reasoning)
This behavior was reportedly reproduced even using DeepSeek's official API, so it wasn't just a quirk of the developer's local quantized setup. It suggests that picking Low to keep costs down can actually make reasoning drag on longer. There are also reports of a bug where reasoning modes don't work correctly through OpenRouter, and since benchmarks have only been published for Max mode, some point out that it's hard to see how the other modes actually perform.
Kota: Bringing Multiple AI Agent CLIs to One "Round Table"
To cut down on the back-and-forth between individual AI coding agents, the open-source project "Kota" has been released, gathering multiple CLIs into a single space. Inspired by the "Knights of the Round Table," it treats humans as members of the team and is designed so that each agent can converse with the others.
It can pull in major CLIs such as Claude Code, Codex, Gemini (Antigravity CLI), Pi, and OpenCode while keeping your existing login settings intact. What makes it distinctive is that agents are given persistent, editable "identities" rather than mere sessions, complete with file-based long-term memory, self-managed skill sets, and worktree-based workspace isolation. It also offers remote operation via a Telegram bot, a drawing board shared between humans and agents, and a cross-project BBS, with support for Kimi Code reportedly in progress.
Parlor v2: A Local Imitation of "GPT-Live" Running on M3 Pro
A community attempt to fully reproduce OpenAI's real-time voice AI "GPT-Live" in an entirely local environment, dubbed "Parlor v2," has been shared. It runs on an M3 Pro.
The developer initially tried fine-tuning Gemma 4 12B into a full-duplex (simultaneous bidirectional) model by grafting on turn-taking timing and speech output, but abandoned the effort after several unsuccessful attempts. For now, they conclude that the classic cascade approach — chaining speech recognition, an LLM, and speech synthesis — remains the realistic option. Open implementations of full-duplex models are still a tall order, and they are taking a wait-and-see stance until a major player releases a model that rivals GPT-Live. The code is available on GitHub.
Authoryze: Guardrails for AI Agent "Payments"
As AI agents increasingly make payments autonomously, "Authoryze," a tool that provides a layer for controlling their spending, has been posted to Show HN. It appears aimed at reining in the risk of agents executing unintended payments through an authorization mechanism. Announced just recently, details are still limited — the scope of the implementation and the supported payment systems will have to wait for future disclosure.
Practical Lessons from Mass-Producing Personal Apps with Claude Code
Finally, some practical insights shared in Japanese. A developer who has released more than 90 mobile apps as a solo creator summarized on Qiita what they learned while delegating an ever-larger share of implementation to AI coding agents like Claude Code. They point out that it's not as simple as "dump everything on the agent and you can mass-produce apps" — there are pitfalls you only discover once you actually run things. Unlike the "skills" mechanism also discussed recently, the post focuses on tips for real-world operation.