Google Significantly Expands Managed Agents on the Gemini API — 3.6 Flash Support and New hooks/triggers
On July 28, Google announced new features for Managed Agents on the Gemini API. The main additions are **support for Gemini 3.6 Flash**, **hooks and triggers** for controlling agent behavior, and reliability enhancements aimed at production use.
Managed Agents is a platform for developers to build production-ready AI agents, and this expansion enables the automation of more complex workflows. Hooks provide a way to insert custom processing before and after agent execution, while triggers launch agents based on external events. Together, they allow more precise control over autonomous agent behavior.
Google has been steadily building out its agent development infrastructure, and the company's focus on ensuring reliability at the API layer is clearly visible.
---
OpenAI Codex / ChatGPT Work Reaches 10 Million Monthly Users — Detailed Report from Latent Space
In a Latent Space interview, Akshay Nathan, OpenAI's lead for core product engineering, revealed that **Codex and ChatGPT Work have reached 10 million monthly active users**.
According to Nathan, Codex MAU as of January 2026 was less than one-tenth of its current level, which works out to more than 10x growth in just six months. The milestone comes less than two weeks after the official ChatGPT Work launch on July 9.
What's particularly notable is that, working from the premise that **"100x more people use code than write it"**, the product targets people outside the programming world. Features such as Sites, OpenClaw, Memory, Subagents, Finance, and No-Code aim to enable no-code software construction.
This shows that the AI coding market is expanding beyond the traditional developer-tools niche to the far broader population of "software users" as a whole.
---
UK Home Office Rejected Asylum Claims Based on AI Hallucinations — Judge Points Out the Issue
It has come to light that the UK Home Office had **rejected asylum claims based on false information generated by AI (hallucinations)**. The problem surfaced when a judge pointed it out.
According to a report by The Guardian, nonexistent information fabricated by an AI system was actually used as grounds for denying applicants. The case has once again highlighted the dangers of using AI output for administrative decisions without human review.
While AI adoption in government brings efficiency gains, it also carries hallucination risks. Especially in decisions like asylum claims that **directly affect the lives and liberty of those involved**, adopting AI output without a verification process could lead to serious human rights violations.
---
Exabase Achieves SOTA on BEAM, the Hardest AI Memory Benchmark — With a Smaller Model
According to a Hacker News post by Exabase founder Johnny, the company's system **M-1** has recorded the highest score on BEAM, widely considered the most difficult AI memory benchmark. The evaluation covers massive contexts of up to **10 million tokens**.
The scores were 76.9% at 100K tokens, 75.0% at 1M tokens, and 68.0% at 10M tokens — beating previous leaders Hindsight (73.4% / 73.9% / 64.1%) and Honcho (63.0% / 63.1% / 40.6%) at every scale.
Particularly noteworthy is that while all previous leaders relied on **Gemini 3 Pro** (a large model), M-1 uses **Gemini 3 Flash** (a smaller model). A scale of 10 million tokens far exceeds any model's context window, so simply stuffing everything into context won't work. The key to the result is said to be a recall mechanism that works at a fundamental level.
---
DeepSeek V4 Flash Hits 32 tok/s on AMD Ryzen AI MAX+ 395
A post on the LocalLLaMA community reported running **DeepSeek V4 Flash and its speculative decoding draft** on a single AMD Ryzen AI MAX+ 395 (128GB unified memory), achieving decode speeds of up to 32 tok/s.
Known as Strix Halo, this chip is AMD's high-end APU that integrates CPU and GPU with 128GB of unified memory. Both the full DeepSeek V4 Flash model and the speculative draft fit within this memory, delivering practical inference speeds.
The possibilities of GPU-free local LLM inference are expanding once again. Particularly for edge devices and privacy-sensitive use cases, this kind of hardware configuration is becoming a compelling option.
---
SWE-rebench Goes Multilingual — Evaluations Across Go / Java / Python / Rust / TS
SWE-rebench announced a major update, expanding its software engineering evaluation — previously Python-only — to five languages: **Go, Java, Python, Rust, and TypeScript**.
Evaluated models include GLM-5.2, DeepSeek-V4 Pro, and Qwen3.6-27B, with multilingual capability tested on real-world bug-fixing tasks. While most coding benchmarks have centered on Python, this effort to visualize performance differences across multiple languages offers valuable practical guidance for model selection.
With per-language strengths and weaknesses becoming clear, the update is expected to provide more accurate data for model adoption decisions in real-world projects.