InfoQ - AI Architecture - News

Cloudflare Introduces Workflows V2 with Deterministic Execution and 50K Concurrent Workflows

Leela Kumili — Fri, 15 May 2026 14:04:00 GMT

Cloudflare introduces Workflows V2, a redesigned distributed workflow orchestration system with deterministic replayable execution, improved observability, and major scaling upgrades, including 50,000 concurrent instances and 2M queued workflows. It supports AI agents, data pipelines, and background processing with improved reliability across distributed systems.

By Leela Kumili

Anthropic Traces Six Weeks of Claude Code Quality Complaints to Three Overlapping Product Changes

Steef-Jan Wiggers — Thu, 14 May 2026 09:16:00 GMT

Anthropic published a postmortem tracing six weeks of Claude Code quality complaints to three overlapping product-layer changes: a reasoning effort downgrade, a caching bug that progressively erased the model's own thinking, and a system prompt verbosity limit that caused a 3% quality drop. The API and model weights were unaffected. All issues were resolved April 20.

By Steef-Jan Wiggers

AWS WorkSpaces Now Lets AI Agents Operate Legacy Desktop Applications without APIs

Steef-Jan Wiggers — Wed, 13 May 2026 07:31:00 GMT

AWS announced that Amazon WorkSpaces can now serve as managed virtual desktops for AI agents in public preview. Agents authenticate through IAM and operate legacy applications via computer vision and input simulation without APIs. Reflex benchmarks show vision agents consume 45x more tokens than API agents.

By Steef-Jan Wiggers

Netflix Introduces ‘Model Lifecycle Graph’ to Scale Enterprise Machine Learning

Matt Foster — Mon, 11 May 2026 07:30:00 GMT

Netflix has developed a graph-based architecture for managing machine learning systems, called the Model Lifecycle Graph. This system maps interconnections between datasets, models, features, and workflows, addressing challenges in scaling ML operations. It enhances discoverability, governance, and component reuse while supporting a self-service approach for engineers and data scientists.

By Matt Foster

How GitHub Is Securing Agentic Workflows in Modern CI CD Systems

Leela Kumili — Fri, 08 May 2026 14:38:00 GMT

GitHub detailed a defense-in-depth security architecture for agentic workflows in CI/CD pipelines, focusing on isolation, constrained execution, and auditability. The design aims to safely integrate autonomous AI agents while mitigating risks like prompt injection, privilege escalation, and unintended actions, using sandboxed environments, restricted permissions, and full execution traceability.

By Leela Kumili

OpenAI Introduces Websocket-Based Execution Mode to Reduce Latency in Agentic Workflows

Leela Kumili — Thu, 07 May 2026 14:48:00 GMT

OpenAI introduces a WebSocket-based execution mode for its Responses API to improve agentic workflow performance in coding agents and real-time AI systems. The update reduces latency by up to 40 percent by replacing HTTP request-response cycles with persistent connections, improving streaming, tool execution, and multi-step orchestration in production-scale AI systems.

By Leela Kumili

Google Announces GKE Agent Sandbox and Hypercluster at Next '26, Positioning Kubernetes as AI Agent

Steef-Jan Wiggers — Thu, 07 May 2026 10:06:00 GMT

Google announced GKE Agent Sandbox and hypercluster at Cloud Next '26. Agent Sandbox uses gVisor kernel isolation for secure agent code execution at 300 sandboxes per second, built as an open-source Kubernetes SIG Apps subproject. It is currently the only native agent sandbox among the three major hyperscalers. Hypercluster manages a million chips from a single control plane.

By Steef-Jan Wiggers

Inside Claude Code Auto Mode: Anthropic’s Autonomous Coding System with Human Approval Gates

Leela Kumili — Tue, 05 May 2026 14:38:00 GMT

Anthropic has introduced auto mode in Claude Code, enabling multi-step software development workflows with reduced manual intervention. The feature combines automated execution with layered safety mechanisms, including input filtering, action evaluation, and two-stage classification, while maintaining human approval checkpoints for sensitive operations.

By Leela Kumili