Blog
Notes and analysis on AI development.
Anthropic sabotage report puts agent monitoring on trial
Anthropic’s Opus 4 sabotage risk report shows why coding agents need audit trails across logs, pull requests, security events, and external review.
OpenAI Sets ChatGPT Retirement Dates for o3 and GPT-4.5
OpenAI will remove GPT-4.5 and o3 from ChatGPT on separate sunset schedules. The API is unchanged, but teams should separate ChatGPT workflows from API model lifecycles.
NSA MCP guidance warns about GitHub scope, WhatsApp leaks, and agent runtime risk
NSA published MCP security design guidance for AI-driven automation, turning tool permissions, tokens, logs, sandboxing, DLP, and scans into deployment requirements.
Command A+ Brings an Open Agent Model to Two H100s
Cohere released Apache 2.0 Command A+, a 218B MoE model with 25B active parameters, private deployment, tool use, vision, and multilingual agent workloads in scope.
SageMaker Skills turn model customization into coding-agent work
AWS opened SageMaker AI model customization to coding agents through Skills, turning SFT, DPO, RLVR, evaluation, and deployment into reviewable notebook workflows.
Google Pay MCP Brings Payment Integration Into IDE Agents
Google Pay and Wallet now expose an MCP server for docs, account status, pass validation, error metrics, and merchant integration workflows.
Sierra localized Agent Studio in four months with AI coding agents
Sierra published a concrete Agent Studio localization case study covering 900+ frontend files, batch scripts, lint loops, and context-window failures.
Lyft AI Assist cuts support-agent development from six months to two weeks
Lyft showed how LangGraph and LangSmith turned customer-support agents into a self-serve platform with routing, state, evals, and prompt CI.
Hugging Face TITO Warns Agentic RL Teams About Token Drift
Hugging Face explains how retokenizing tool-using agent rollouts can break gradients, and proposes TITO as a safer training-loop rule.
IBM Agentic CLEAR tracks agent failures across three levels
IBM Research released the Agentic CLEAR paper and open source tool for analyzing agent traces at the system, trace, and node levels.
Claude self-hosted sandboxes set new rules for private MCP access
Anthropic added self-hosted sandboxes and MCP tunnels to Claude Managed Agents, shifting tool execution and private tool access into enterprise-controlled boundaries.
Workday ASOR and Gemini move HR agents to the approval line
Workday and Google Cloud connected Sana to Gemini Enterprise. For HR and finance agents, approval chains, permissions, and data boundaries matter more than the model.