AI Agents Enter the Era of "Scheduled Execution"—The Big Three Launch Simultaneously
June 2026 is seeing a major seismic shift in the AI industry. OpenAI, Anthropic, and Google—the three giants of AI platforms—have successively released "scheduled-execution AI agents."
OpenAI**: Scheduled Tasks for ChatGPT (released June 17)
Anthropic**: Managed Agents for Claude (cron-based, CLI/auth service integration, vault-managed environment variables)
Google**: scheduling features for Gemini
Until now, AI agents have basically worked "only when the user issues an instruction." From here on, the premise flips: they will "autonomously execute tasks at set times." Design patterns for scheduling AI like cron jobs look set to become essential knowledge for developers.
DeepGate's Edge AI Compiler Outperforms Google's
DeepGate, a company developing AI compilers for edge devices, has announced that its compiler surpasses Google's toolchain and the toolchains of various hardware vendors in terms of performance.
Alongside this, the company has also unveiled an effort to automate model design for edge AI using neural architecture search (NAS). Edge devices operate under severe resource constraints, making model optimization critically important—and the ability to automate this step with NAS could accelerate the adoption of edge AI.
AI Startup Fraud—Australian Entrepreneur Sentenced to 9 Years in Prison
An Australian entrepreneur who billed the venture as an "AI startup" despite having no actual AI, deceiving investors and falsifying revenue, has been sentenced to nine years in prison.
As the AI bubble heats up, a situation has emerged where merely claiming to be AI is enough to attract investment. This verdict stands as a symbolic event signaling stronger law enforcement against false AI claims. For investors, it also underscored, once again, the importance of technical due diligence.
A $1800 Local LLM Build Runs 262K Context at 55 tok/s
A build making the rounds in the LocalLLaMA community connects four RTX 5060 Ti (16GB) GPUs via P2P and runs Qwen 3.6-27B (FP8) with a 262K context and BF16 KV cache at 55 tok/s. The total GPU cost is just $1800 (about ¥270,000).
The fact that individuals can now build local environments capable of handling enormous context windows without relying on cloud APIs carries major implications for both privacy and cost. The evolution of consumer GPUs is rapidly expanding what local LLMs can do.
"Agent Device" Lets AI Agents Operate iOS Apps
"Agent Device," released by Callstack Incubator, is a CLI tool for AI agents that can operate iOS, Android, TV, and desktop apps. Built on the accessibility tree, it accesses UI elements via references (e.g., @e1), allowing coding agents like Claude Code and Codex to "see and touch" real devices.
Beyond automated mobile app testing and accessibility audits, this could open up new use cases such as AI autonomously operating mobile apps in the future.
Roughly 400K Sessions Back Up "Humans Plan, AI Executes"
In "Agentic coding and persistent repositories," a study published by Anthropic, an analysis of approximately 400,000 sessions reveals that experienced Claude Code users follow a pattern of "humans handling design and judgment while delegating execution to AI."
Specifically, the effective division of labor is this: verify irreversible operations with your own eyes before executing them, and leave code details to the AI. As AI coding tools continue to spread, what draws attention is that the study demonstrates an appropriate division of roles between humans and AI based on empirical data.
---