// edition · 2026-05-14
27 stories on AI dev tools, agents, and the coding stack — curated from the day's RSS haul by Agentic Dev's pipeline.
Anthropic's Claude Code Ultraplan, described as a research preview, separates the planning phase from code execution by offloading plan drafting to a cloud session, allowing users to review and comment on plans in a browser before execution. The feature requires Claude Code v2.1.91 or later and i...
Anthropic announced that starting June 15, 2026, Claude Agent SDK usage — including `claude -p` automation, Claude Code GitHub Actions, and third-party SDK-authenticated apps — will be billed against a separate monthly credit rather than subscription rate limits, with Max 20x subscribers receivin...
A developer building an AI agent found that passing data as raw JSON instead of plain prose used 2.6x more tokens, resulting in roughly 2.5x higher API costs per call. The difference stems from how BPE tokenization handles JSON structural characters like braces, quotes, and colons as separate tok...
Anthropic released an "agent view" dashboard for Claude Code that lets developers monitor and manage multiple AI coding sessions from a single CLI interface, showing session status and enabling inline replies. Developer reactions are mixed, with some welcoming the centralized view while others ar...
Several AI coding agents compete with Anthropic's Claude Code, including OpenAI's Codex, which offers built-in browser access and cloud environments, and openCode, an open-source alternative. Most offer free tiers, and the tools vary in form factor between CLI, TUI, and full applications.
An analysis in The New Stack argues that AI coding agent performance depends more on surrounding scaffolding — prompts, tools, and feedback loops — than model selection, citing data showing the same model moved from rank 30 to rank 5 on Terminal Bench 2.0 with a different harness. The piece conte...
A Celery worker running Claude Code CLI as a subprocess was intermittently failing with a misleading "Control request timeout: initialize" error, which turned out to be the Linux kernel OOM killer terminating the CLI process mid-startup. The fix was routing the task to a dedicated ECS Fargate que...
OpenAI built a secure sandbox for its Codex coding agent on Windows, implementing controlled file access and network restrictions to allow safe execution of automated coding tasks.
An engineering team at Port built a centralized library of AI coding assistant "skills" — Markdown configuration files defining company standards — after discovering each engineer was running different, untracked local configurations. The library is stored in version control, allowing engineers t...
Model routing directs AI prompts to different models based on complexity, cost, and latency, rather than using a single model for all queries. Cloud providers including Microsoft Azure AI Foundry and AWS Bedrock have released built-in routing tools trained on datasets spanning question answering,...
Anthropic announced Claude for Small Business, a version or plan of its Claude AI assistant targeted at small business users.
A technical analysis describes using Microsoft Intune's Security Copilot integration to automate endpoint remediation at enterprise scale, converting endpoint signals into AI-driven, governed remediation actions. The piece applies a proprietary methodology called the Rahsi Framework™ to evaluate ...
Anthropic launched Claude for Small Business on May 13, 2026, offering 15 agentic workflows, 15 reusable skills, and connectors to eight platforms including QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, Microsoft 365, and Slack.
A developer released "tsk," an open-source local MCP server written in Go that proxies API calls for LLM agents, injecting credentials at runtime without exposing them to the model. It enforces an allowlist via rules.yaml, scrubs sensitive data from responses, applies per-tool rate limits, and lo...
A developer comparison of Claude (Anthropic) and ChatGPT (OpenAI) in 2026 found Claude Opus 4.6 scores 80.8% on SWE-bench Verified versus GPT-5.4's roughly 80%, and 91.3% on GPQA Diamond reasoning benchmarks. Both services cost $20/month; Claude was rated stronger for long-context coding and regu...
A developer released Network-AI, an open-source coordination layer for multi-agent AI systems that uses atomic propose-validate-commit cycles to prevent concurrent state overwrites. The project claims support for 14 frameworks including LangChain, AutoGen, CrewAI, and the Model Context Protocol.
A journalist tested GPT-5.5 Instant against GPT-5.2 after OpenAI replaced its default ChatGPT model, finding the conciseness claim did not hold up — GPT-5.2 produced shorter answers in all three test cases — while GPT-5.5 showed reduced hallucinations on factual queries.
A developer investigated why Claude produces verbose responses by analyzing RLHF training mechanics, arguing that human annotators in the reward model training phase tend to prefer longer responses, which reinforces verbosity as a learned prior. The author built a reward model simulation using An...
Notion launched a developer platform that allows teams to connect AI agents, external data sources, and custom code directly into their Notion workspace.
A supply chain attack on the TanStack npm package, dubbed "Mini Shai-Hulud," compromised OpenAI signing certificates and systems. OpenAI is requiring macOS users to update its apps by June 12, 2026, as part of its remediation response.
GitLab's 2025 Global DevSecOps Survey found developers spend about 15% of their time writing code, with the remainder on planning, reviewing, testing, and coordination. Enterprises are increasingly adopting multi-model AI strategies, routing tasks to different models based on cost, speed, and qua...
Anthropic launched Claude Finance on May 5, a bundle of 10 agent templates covering five financial services domains — investment banking, equity research, private equity, wealth management, and financial analysis. The package includes Microsoft 365 integration and ships via Claude Cowork and Clau...
Cat Wu, Anthropic's head of product for Claude Code and Cowork, said the next major step for AI is proactivity — systems that anticipate user needs before users are aware of them.
Anthropic now has more business customers than OpenAI among companies tracked by fintech firm Ramp, with 34.4% of Ramp's clients paying for Anthropic services versus 32.3% for OpenAI.
Adaption released AutoScientist, a tool that automates the fine-tuning process for AI models, allowing them to adapt to specific capabilities without manual intervention.
MinIO launched MemKV, a petabyte-scale flash-based context memory store for AI inference workloads, accessed over 800 Gigabit Ethernet RDMA. The company claims it reduces GPU recompute by retaining context across GPU clusters, achieving 95% better GPU utilization and roughly 50% lower cost per to...
Red Hat announced a dedicated AI skills repository at its Summit in Atlanta, offering "skill packs" that layer agent capabilities on top of Red Hat Enterprise Linux, OpenShift, and Ansible. The company's Ask Red Hat chatbot, now on its Customer Support Portal, was trained on over 20 years of Red ...