Archive
Adding capability made the agent less capable
LangChain benchmarked NVIDIA's NeMo Switchyard router across multi-step agentic tasks, and the spend came back wildly lopsided
The winning setup used less than half the tokens and finished more tasks
Agent orchestration is quietly becoming a managed service
RAG's top-k retrieval has a black box problem, and it isn't fixed by a bigger model
Most agent security still stops at the tool call.
Self-improving agents can poison their own skill library, and nobody is checking before it happens
Search built for agents doesn't look like search built for humans
LangSmith just turned every coding agent into one shared trace
Most coding agents run setup instructions without asking a single question
Two labs shipped agent memory frameworks in June 2026. They disagree on one thing
Most teams sandbox agent code in containers. LangChain argues you do not need to
Anthropic shipped Claude Sonnet 5. The part that matters is not the benchmark
Codex says 20 percent of its users aren't developers. That number is the whole story
The reliable agents - all do the same boring thing: they check their own work
An AI agent reverse-engineered malware that most antivirus tools missed
Every code assistant you use predicts one token at a time. Google shipped one that does not
Anthropic shipped two models on June 9. The one you cannot use is the more interesting
Two vendors named the same thing. The agent isn't the model. It's the harness
Asking your AI to summarize a web page is now an attack surface
Everyone is still trying to prompt their agents into reliability. The tooling quietly moved the other way last week
Anthropic shipped hundreds of parallel agents last week. The part that matters is the one that checks them
Work today feels less like building and more like operating a system
I use AI in almost everything I do. But I don’t rely on it