AI agents gain memory, safety; regulation advances

Anthropic / Claude ecosystem

Anthropic adds persistent memory to Claude Managed Agents in public beta

Anthropic has launched a persistent memory layer for Claude Managed Agents, enabling long-running AI agents to retain learning across sessions. Early adopters have reported a 97% error reduction and a 30% speedup in production workflows with this feature, which also includes workspace-scoped controls and audit trails.

Frontier model providers

OpenAI Agents SDK Update Adds Sandbox Execution for Safer, Long-Running Enterprise AI Agents

OpenAI's Agents SDK now features secure sandboxed execution for long-running enterprise AI agents. This update includes audit tracing, approval gates, and isolated file/command workflows to mitigate unauthorized system access risks and enable safer autonomous operations.

Meta AI Releases Sapiens2: A High-Resolution Human-Centric Vision Model for Pose, Segmentation, Normals, Pointmap, and Albedo

Meta AI has released Sapiens2, a new foundation model for human-centric vision that scales to 5B parameters and 4K resolution. It combines masked autoencoder and contrastive learning pretraining with 10 times more task-specific supervision across various human image analysis tasks.

DeepSeek 全系 API 輸入緩存降價至首發價十分之一 中國 | Techritual 香港

DeepSeek has significantly reduced its API input cache pricing to 1/10 of its launch price, effectively undercutting GPT-5.5 Pro by 700 times. This aggressive pricing move is expected to force an industry-wide pricing realignment for large language models.

OpenAI Agents SDK April 2026: Sandbox Agents, Harness Architecture, Any LLM

The OpenAI Agents SDK April 2026 update introduces sandbox agents with persistent isolated workspaces and harness-compute separation for enterprise auditability. It also adds compatibility with over 100 LLMs via the Chat Completions API, enabling safer and more capable autonomous coding agents.

AI developer tooling & infrastructure

Cursor 3 'Glass' Replaced Composer with an Agents Window

Cursor 3 'Glass' has replaced its Composer side-pane with a full-screen Agents Window, allowing users to orchestrate multiple AI agents in parallel. This feature supports execution across local worktrees, cloud sandboxes, and remote SSH targets.

GitHub Copilot code review will start consuming GitHub Actions minutes on June 1, 2026

GitHub has announced that Copilot code reviews on private repositories will begin consuming GitHub Actions minutes starting June 1, 2026. This change will require customers to factor in new billing charges for their AI-assisted code review processes.

Sachitrafa Releases YourMemory Agent Layer

Sachitrafa has launched YourMemory Agent Layer, which provides AI agents with persistent, human-like memory without external service dependencies. This local, zero-infrastructure solution aims to reduce repetitive interactions and improve contextual understanding.

Llama 4 API Access: Complete Developer Guide (Scout, Maverick, ofox)

Meta's Llama 4 Scout model now offers a 10M-token context window at $0.08/M input tokens. This makes long-document analysis economically viable for developers and enterprises without vendor lock-in, broadening access to advanced generative AI capabilities.

Cloud & platform providers

Extend AI voice support with real‑time voice agents in Copilot Studio | Microsoft Copilot Blog

Microsoft has announced the general availability of real-time voice agents in Copilot Studio, enabling low-latency, interruptible speech-to-speech conversations with real-time reasoning. This feature is designed to scale enterprise customer service operations.

Cloudflare Boosts AI Agent Governance; Claude Model Choice & Advanced NLP

Cloudflare has released enterprise Model Context Protocol (MCP) governance features for AI agent orchestration, integrating Cloudflare Access authentication and Code Mode. This addresses secure production deployment patterns for multi-cloud AI systems, supporting models like Claude 4.7.

Structured error responses for Cloudflare 5xx errors · Changelog

Cloudflare now returns structured JSON and Markdown responses for 5xx errors, complying with RFC 9457. This enhancement enables better error attribution and retry handling for API clients interacting with Cloudflare services.

AI policy, regulation & governance

Lieu and Obernolte introduce consolidated AI bill package

A bipartisan bill, the 'American Leadership in AI Act,' has been introduced in the US Congress, consolidating over 20 previous AI regulatory proposals. The comprehensive package covers AI standards, research, governance, worker protection, and deepfake deterrence.

EU Regulators Advise Google on AI Access and Compliance Measures

EU regulators have issued binding guidance requiring Google to open Android AI integration and search data access to competitors. This decision, under the Digital Markets Act, targets compliance by the end of July, with potential fines up to 10% of global sales for violations.

DNV and industry leaders collaborate to strengthen Norway’s competitiveness with new industrial AI safety framework

Norway has launched a first-of-its-kind cross-sector industrial AI safety regulatory sandbox, bringing together telecommunications, finance, energy, maritime, and defense. Led by DNV, this initiative aims to establish safety standards for autonomous systems and AI deployment.

Industry & market moves

Google DeepMind Launches Inaugural AI Campus in South Korea Through Strategic Partnership

Google DeepMind has launched its inaugural AI campus in South Korea, integrating advanced models like AlphaEvolve and AlphaGenome with local academic institutions. This initiative aims to accelerate scientific breakthroughs, AI safety, and workforce development, aligning with Korea's K-Moonshot initiative.

China blocks Meta’s $2B Manus deal after months-long probe

China's NDRC has blocked Meta's $2 billion acquisition of Singapore-based AI startup Manus after a months-long probe, citing national security concerns. This marks a significant intervention in a cross-border tech deal and a setback for Meta's AI agents ambitions.

Singtel Digital InfraCo’s RE:AI and Mistral AI partner to strengthen Singapore’s industry AI capabilities

RE:AI, Singtel Digital InfraCo's unit, and Mistral AI have formalized a strategic partnership to build sovereign AI infrastructure and secure multi-tenant GPU services in Singapore. This collaboration aligns with Singapore's National AI Strategy 2.0.

TCS expands Google Cloud partnership for autonomous AI

Tata Consultancy Services (TCS) and Google Cloud have expanded their partnership to help enterprises transition from AI pilots to autonomous, AI-native operating models. This includes four new products covering data, industrial automation, and security, enabled by Google SecOps.

AI product & feature launches

Meta unifies all accounts under one roof

Meta is unifying account management across Facebook, Instagram, Threads, and its AI services under a single 'Meta Account.' This new system supports passwordless authentication and includes an expanded Business Assistant for advertisers.

Turning customer experience into a growth engine - Microsoft Dynamics 365 Blog

Microsoft announced the general availability of agentic customer experience capabilities, including real-time voice agents and AI-driven agents across Dynamics 365 Contact Center, Sales, and Customer Insights. This aims to automate routine tasks and improve customer outcomes.

ASAPP Launches Multiple AI Agents Within CXP to Advance

ASAPP's Customer Experience Platform (CXP) now integrates five purpose-built AI agents (Discovery, Developer, Simulation, Insights, Optimization). These agents are designed to orchestrate end-to-end enterprise customer service, addressing deployment, governance, and consistency at scale.

BeyondTrust Launches PathfinderAI to Power AI-Native Identity Security Operations

BeyondTrust has introduced PathfinderAI, featuring natural-language query capabilities for identity risk investigation and a new Model Context Protocol (MCP) Server. This MCP Server enables external AI agents (like Copilot, OpenAI, Claude) to securely access identity security intelligence.

Identity Digital Launches Neutral, DNS-Anchored Identity

Identity Digital has introduced DNSid, a persistent DNS-anchored identity standard specifically designed for AI agents. This new standard binds autonomous systems to verifiable ownership across various platforms and organizations.

Research with immediate practical relevance

Google DeepMind Paper Argues LLMs Will Never Be Conscious

A Google DeepMind scientist has published a philosophical argument, 'The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness,' asserting that large language models (LLMs) can never achieve consciousness. This contradicts some AGI narratives from AI company CEOs.

New AI framework autonomously optimizes training data, architectures and algorithms — outperforming human baselines

The Generative Artificial Intelligence Research Lab (SII-GAIR) has developed ASI-EVOLVE, an AI framework that autonomously optimizes AI training data, architectures, and algorithms. It discovers novel designs that outperform human baselines by more than 18 points on MMLU and other benchmarks.