Blog
Notes and analysis on AI development.
Copilot Studio Now Clicks Apps Without APIs
Microsoft Copilot Studio computer use GA moves UI automation agents from demos into enterprise deployment, audit, and governance.
One API Call Now Boots Linux, the Line Google Drew with Gemini Agents
Google Managed Agents extends the Gemini API from model calls into sandboxed execution, shifting where agent infrastructure begins.
When 65% of Teams Treat the IDE as Optional
Gartner and OpenAI show enterprise coding agents shifting from model benchmarks toward governance, cost control, and deployment architecture.
The 7,000-return loop behind Codex self-improving agents
OpenAI Tax AI shows why production traces, eval sets, and practitioner feedback matter more than agent automation alone.
From Inbox to DCF, Why Codex Is Moving Beyond Code
OpenAI Codex use cases now span inboxes, data, finance, QA, app automation, and collaboration, a sign that coding agents are becoming work agents.
Agent memory moved into files, and AMP shows a different path from OpenAI
OpenAI Agents SDK memory and the AMP v0.1 draft turn long-term agent memory into files, Git history, MCP resources, and auditable state.
Cohere Command A+ Puts a 218B Agent Model on Two H100s
Cohere Command A+ pushes the private agent model race with Apache 2.0 weights, a 218B MoE design, and a two-H100 W4A4 deployment path.
NVIDIA Vera arrives, and agent infrastructure gets a CPU bottleneck
NVIDIA’s first Vera CPU deliveries show that agent infrastructure bottlenecks are spreading from GPU inference into CPU orchestration.
CLAUDE.md Became an Attack Surface, TrapDoor's Invisible Instructions
TrapDoor combines malicious npm, PyPI, and Crates.io packages with poisoned AI coding instruction files.
Mistral Picks a Narrow Path Into Industrial Agents With Emmi AI
Mistral AI’s acquisition of Emmi AI shows the LLM race moving into physics simulation, CAD/CAE workflows, and industrial R&D agents.
Claude Containment Shows Agent Security Is Now Blast Radius Design
Anthropic’s Claude containment write-up shows agent security moving from prompt defenses toward environment isolation, scoped tokens, and blast-radius control.
CUDA 13.3 adds a new tuning lever for LLM inference
NVIDIA CUDA 13.3 targets the lower layers of LLM inference cost with CompileIQ compiler tuning and CUDA Python 1.0.