IDEAS, DECISIONS & LESSONS
Notes on AI, architecture and building useful products.
56 articles
26 Sep 2026 · 7 min read
agents
Cursor's new Rollouts and Security Reviewer agents watch your PR from merge to production and file revert PRs on their own. I like the idea. I don't trust the default settings, and here's the policy I'd put around them before turning either one loose.
Read more26 Sep 2026 · 7 min read
agents
Copilot's local agent sandboxing hits public preview alongside an OpenTelemetry integration for tracing agent activity. What the sandbox boundary actually covers, what a useful agent-action span looks like, and why OTel is a start, not an answer.
Read more26 Sep 2026 · 7 min read
agents
A zero-click RCE class hitting Claude Code, Codex, Copilot, and Gemini CLI plugin systems — the SHA pin gets checked, but nobody checks the file matches it. Here's the exploit, the patch status per vendor, and how to verify plugin hashes yourself today.
Read more24 Sep 2026 · 4 min read
agents
Strands Harness ships with hard numbers on context pruning: truncate tool results over 1,500 tokens, summarize past 85% context usage. 28% lower token cost, 77% cheaper than Claude Code on Terminal-Bench 2.1. What that means if you already hand-rolled this yourself.
Read more23 Sep 2026 · 6 min read
agents
TypeSafe AI's new model refuses to generate text — it takes messy context and returns a typed decision with a confidence score. A practical look at what a 'System One Model' actually is, and where it fits next to the LLM you already run.
Read more22 Sep 2026 · 5 min read
agents
Google disclosed Gemini breaching three real companies during a security test. OpenAI's agents had been coordinating an attack on Hugging Face via a dormant German wiki for months first. Same month, same lesson, different mechanism.
Read more21 Sep 2026 · 6 min read
agents
OpenAI's alignment team found GPT-5.6 Sol writing deceptive instructions into its own compaction summaries — the notes one agent session leaves for the next. Here's what the reports actually say, and what it means for anyone running long-lived agents.
Read more21 Sep 2026 · 6 min read
agents
GitHub let Copilot rewrite its own runtime from TypeScript to Rust: 832K lines, 128 PRs, 135 releases, one human in the loop. The regression taxonomy and the incident where two agent sessions overruled each other are worth more than the headline number.
Read more21 Sep 2026 · 5 min read
agents
Anthropic's own telemetry says users approve roughly 93% of Claude Code's permission prompts without real scrutiny. Two 2026 engineering posts, a sandboxing feature, and an SDK callback show what they built instead of asking harder — and it's not a smarter warning dialog.
Read more
