Weekly | AI New Product & Feature News (June 7 - June 13, 2026)
Update Date: 2026/6/14
Executive Summary
This week's new AI products and features focused on building the foundation to move AI agents from experimental stages to production. OpenAI, AWS, GitHub, and Microsoft have strengthened long-running execution, interactive debugging, CI integration, and cloud agent supervision to advance safe autonomous execution in development environments. Apple and Google are integrating AI into OS, browsers, and analytics platforms, accelerating its penetration into daily tasks and personal use. Domestically, Ricoh, freee, and LINE Yahoo are expanding business-specific and consumer-oriented AI, while the importance of auditing, retention, data residency, and evidence tracking is increasing significantly.
*The created article content was input into Gamma to automatically generate slides. If you find slides easier to view, please take a look here.
1. Expansion of AI Agent Execution Infrastructure
1-1. OpenAI: Strengthening long-running cloud execution for Codex with the acquisition of Ona
Source URL: OpenAI "OpenAI to acquire Ona" / OpenAI Help Center "ChatGPT Release Notes"
OpenAI announced several updates to Codex in its release notes dated June 11. A Developer mode has been added for Chrome and the browser with built-in Codex, enabling controlled access to the Chrome DevTools protocol, such as JavaScript profiling, inspection of console and network traffic, and DOM verification. Additionally, an automatic generation feature for AGENTS.md project instruction files via the /init command has been introduced.
1-2. Amazon Bedrock AgentCore Runtime: Adding interactive shells to running agents
Source URL: AWS What’s New "Amazon Bedrock AgentCore Runtime"
AWS has added interactive shells to the Amazon Bedrock AgentCore Runtime, allowing access to running agent sessions via WebSocket. Developers can check files, investigate environment states, execute ad-hoc commands, and perform debugging on isolated microVMs. This update enhances observability and operability for running sessions when deploying coding agents like Claude Code, OpenAI Codex, and Amazon Kiro into production, and is important as a foundation for safely supervising agents.
1-3. Microsoft: Expanding business agent operations with Copilot Notebooks and Power Platform
Source URL: Microsoft 365 Copilot Blog "What’s new in Notebooks" / Microsoft Power Platform Blog
Microsoft has started offering Copilot Notebooks to Copilot Chat users, allowing them to add files to ask cross-cutting questions within the notebook and use learning tools like mind maps. In Power Platform, closed-loop learning has been introduced for the Power Apps MCP server. A mechanism has been added where corrections made by users in the agent feed are accumulated and reused as structured memory, allowing agents to autonomously learn and improve organization-specific business patterns.
2. Full-scale Development AI and Coding Agents
2-1. GitHub: Implementing "Thinking CI" on Actions with Agentic Workflows
Source URL: GitHub Changelog "GitHub Agentic Workflows" / GitHub Changelog "GITHUB_TOKEN support"
GitHub Agentic Workflows is now in Public Preview, allowing inference-based tasks such as issue triage, CI failure analysis, and documentation updates to be executed on GitHub Actions. When automation content is defined in natural language Markdown, it is converted into standard Actions YAML, leveraging existing runners and permission management. At the same time, with GITHUB_TOKEN support, long-lived PATs are no longer required, and features such as read-only permissions, sandboxing, and output verification are included, establishing a foundation for safely operating agents in development environments.
2-2. GitHub Copilot: Strengthening cloud agent supervision and security verification
Source URL: GitHub Changelog "Copilot Chat now sees your agent sessions" / GitHub Changelog "Security validation for third-party coding agents"
GitHub has added a feature that allows Copilot Chat to reference cloud agent work sessions, enabling status checks in progress and follow-up questions after completion. Furthermore, it has generally released a feature to apply automatic verification via CodeQL, Advisory Database, and secret scanning to code generated by third-party coding agents such as Claude and OpenAI Codex. As the premise of AI writing code becomes more widespread, the trend of humans supervising through conversation and the platform side verifying safety is strengthening.
2-3. GitLab & Microsoft Foundry: Building infrastructure to lower the barrier to entry for agent development
Source URL: GitLab Press Release "GitLab Orbit" / Microsoft Tech Community "Foundry Hosted Agents"
GitLab announced "GitLab Orbit," a context graph that spans code, work items, pipelines, deployments, and production signals, in public beta. It claims to speed up agent responses by up to 11 times and reduce token consumption by a factor of 4.5. Meanwhile, Microsoft added a preview feature to Foundry Hosted Agents that allows Python source code to be uploaded and deployed as a ZIP file, providing a new deployment path that does not require containers.
3. Progress in Model/OS Integration by Apple, Google, and Anthropic
3-1. Apple: Siri AI and Apple Intelligence for Cross-OS Personal Agents
Source URL: Apple Newsroom 'Siri AI' / Apple Newsroom Japan
At WWDC26, Apple completely revamped Siri as 'Siri AI,' evolving it into a conversational assistant that integrates screen understanding, personal context, web information, and cross-app actions. It spans multiple devices including iPhone, iPad, Mac, and Vision Pro, assisting with answering questions and executing tasks based on information from Mail, Messages, Photos, Spotlight, and more. With a design combining on-device models and Private Cloud Compute, the direction of embedding AI as a personal agent resident across the entire OS has become clear.
3-2. Apple Xcode 27 & Foundation Models: Integrating AI Agents into Development Environments
Source URL: Apple Newsroom 'New intelligence frameworks and advanced tools' / Google Blog 'Gemini models to Apple developers'
In Xcode 27, Apple integrated models and agents from Anthropic, Google, and OpenAI into the development flow, demonstrating support for conversational planning, test execution, preview confirmation, MCP, and the Agent Client Protocol. The Foundation Models framework has also been expanded, allowing access to on-device models, server models, image input, and custom skills via Swift APIs. Furthermore, Google announced Gemini in Xcode, advancing the trend of external AI models naturally entering the Apple development environment.
3-3. Anthropic, Google, and Cohere: Model Competition Shifts to Long-term Tasks and Development Use Cases
Source URL: Anthropic 'Claude Fable 5 / Mythos 5' / Google Blog 'DiffusionGemma' / Cohere Release Notes
Anthropic announced Claude Fable 5 and the specialized Claude Mythos 5, emphasizing long-term knowledge work, software development, visual understanding, and cyber defense. Google released the text-diffusion open model DiffusionGemma, highlighting speed through parallel generation. Cohere also released the code-specialized MoE 'North Mini Code,' presenting a lightweight model for developers that supports long contexts. The model competition is expanding from mere overall performance to long-term tasks, development execution, and ease of local/enterprise deployment.
4. Strengthening Connections with Business SaaS and Enterprise Data
4-1. Salesforce: Accelerating Domestic Agent Adoption with AgentExchange and FDE Partner Network
Source URL: Salesforce 'AgentExchange' / Salesforce 'FDE Partner Network'
Salesforce Japan has launched AgentExchange in Japan, integrating AppExchange, Slack Marketplace, and the Agentforce ecosystem. Companies can discover, purchase, activate, and manage AI agents, sub-agents, tools, MCP servers, and business apps from a single marketplace. Additionally, the Forward Deployed Engineering (FDE) Partner Network, which supports Agentforce production deployment, has been fully launched domestically, establishing a system to support the transition from pilot to production with nine initial partners.
4-2. Google: AI-Powered Daily Tasks and Analytics with Gemini in Chrome, Looker, and BigQuery
Source URL: Google Blog 'Gemini in Chrome' / Google Cloud Blog 'dashboard agents in Looker' / Google Cloud Release Notes
Google has expanded the availability of Gemini in Chrome to Latin America, Africa, the Middle East, and other regions, enabling page summarization, cross-tab comparison, and integration with Gmail, Calendar, Maps, and YouTube directly from the browser. In Looker, 'dashboard agents' are now in Preview, allowing users to embed conversational agents into dashboards to perform deep dives into metrics and additional analysis using natural language. Furthermore, in BigQuery, Gemini Cloud Assist has added data lineage analysis and query scheduling features in Preview, strengthening the trend of AI consistently supporting everything from investigation and analysis to operational settings.
4-3. OpenAI, Oracle, and AWS: The Trend of Using AI Models Within Existing Cloud Contracts
Source URL: OpenAI 'OpenAI on Oracle Cloud' / AWS What’s New 'OpenAI GPT on Amazon Bedrock'
OpenAI and Oracle have introduced a new procurement path that allows OCI customers to use existing Oracle cloud commitments to access OpenAI models and Codex. AWS has also expanded the regions for using OpenAI GPT-5.4/GPT-5.5 on Amazon Bedrock, making it easier for enterprises to handle high-performance models within their existing cloud permission management, auditing, and operational rules. The focus of AI adoption is shifting from which model is strongest to how it can be integrated into existing contracts, data residency, auditing, and operational infrastructure.
5. Expansion of Domestic AI New Products and Business-Specialized Agents
5-1. Ricoh: Supporting In-House AI Utilization with Dify App Sharing Infrastructure and Japanese LMMs
Source URL: Ricoh 'Dify App Marketplace' / Ricoh 'Qwen3.6-Ricoh'
Ricoh has begun internal operation of a community-based Dify app marketplace utilizing the generative AI app development platform 'Dify,' with plans to offer it to customers starting this summer. It features a mechanism to share, search, and download useful apps from the approximately 9,300 apps operated within the group, and includes an AI agent for hearing requirements that automatically generates not only the app itself but also user manuals and implementation procedures. Meanwhile, based on Alibaba Cloud's Qwen3.6-27B and Qwen3.5-9B, they have developed the multimodal large language models 'Qwen3.6-Ricoh-27B-20260522' and 'Qwen3.5-Ricoh-9B-20260522' with enhanced Japanese reasoning performance, providing quantized versions that can be operated in on-premises environments to support the use of highly confidential corporate documents.
5-2. freee, FreakOut, and Benri.ai: AI Agents Penetrating Back Office, Advertising, and Sales
Source URL: freee-mcp Press Release / FreakOut 'HAWK' / Benri.ai Press Release
The freee MCP server 'freee-mcp' now supports freee IT management, allowing AI agents to execute tasks such as SaaS account issuance, permission changes, stopping access for departing employees, and IT asset inventory. FreakOut has launched the SNS advertising operation AI agent 'HAWK,' which automates quoting, setup, delivery monitoring, and reporting. Benri.ai has also officially released an AI mode that advances sales preparation through conversation, indicating that agent utilization in domestic business settings is spreading.
5-3. LINE Yahoo and SwitchBot: AI for Consumers Also Expanding to Memory, Generation, and Home Appliance Control
Source URL: LINE Yahoo 'Agent i' / SwitchBot 'KATA Friends'
LINE Yahoo has added image generation, image editing, conversation tone selection, and memory functions to its AI agent 'Agent i,' expanding daily-use AI from search and conversation to memory, generation, and domain-specific tasks. SwitchBot has also added the 'Help' home appliance control function to its AI partner robot 'KATA Friends,' enabling operation and status checks for lighting, air purifiers, air conditioners, robot vacuums, and more through conversation alone. AI is evolving not only in the business domain but also as a lifestyle interface that connects home devices.
6. Increasing Importance of AI Governance, Auditing, and Security
6-1. Google Vault & Security Command Center: Bringing AI Usage into Legal Retention and Defense Scopes
Source URL: Google Workspace Updates "Google Vault Gemini app retention" / Google Cloud Release Notes
Google Workspace's Google Vault now supports retention rules and legal holds for Gemini apps, allowing generative AI conversations to be treated as subjects for eDiscovery and legal holds. If a user deletes a conversation, it can still be retained and reviewed by Vault administrators if a legal hold is active. Additionally, AI Protection in Security Command Center has added support for EU data residency, making it easier to advance the monitoring and defense of AI usage even in regulated industries. As AI adoption spreads, the management of conversation logs, retention, auditing, and data location is becoming increasingly critical.
6-2. LlamaIndex, Megaport, and UMP: Grounding Traceability and Connection Standards in the Agent Era
Source URL: LlamaIndex Newsletter / Megaport MCP Server / Universal Memory Protocol
LlamaIndex has added word, line, and cell-level granular coordinates to LlamaParse, strengthening the audit trail that allows AI to return to the exact location within a document from extracted values. Megaport now allows AI assistants to query network operation information in natural language via an MCP Server, and UMP is proposing a common protocol for carrying memory between agents. In an era where agents handle multiple tools and data, grounding traceability, connection standards, and memory portability are becoming essential foundational themes.
Comprehensive Analysis
The major shift visible from this week's trends is that AI is evolving from standalone chats or auxiliary tools into execution entities that reside in business systems, development platforms, operating systems, cloud contracts, and home devices. In particular, updates from GitHub and AWS assume an operational model where agents write code and maintain sessions, while humans supervise and correct them along the way. Meanwhile, moves to make AI output grounding and conversation logs subject to audits, such as those by Google Vault and LlamaIndex, have also intensified. Moving forward, operational design—including permission management, verification, memory, auditing, and integration with existing workflows—will determine competitiveness more than the mere "introduction of high-performance models."
Points to Watch in the Future
AI agents are moving from the stage of "delegating tasks" to "humans observing and intervening during execution," making supervision UIs and permission design the deciding factors for successful adoption.
Due to moves by Apple, Google, and Microsoft, AI is being integrated into operating systems, browsers, and development environments rather than just individual apps, intensifying the battle for control over user touchpoints.
For corporate AI adoption, selection criteria will increasingly focus not just on model performance, but on compliance with existing cloud contracts, data residency, audit logs, and legal retention.
In the domestic market, the use of Dify, MCP, and task-specific agents is advancing, and the ability to port site-specific procedures and know-how into AI will become a differentiator.
As agents cross over multiple tools, safety foundations such as output grounding traceability, memory portability, and third-party code verification will become indispensable.
