https://learning-notes-dz2.pages.dev/categories/agentic/ Recent content in Agentic on My Learning Notes Hugo -- 0.124.0 en Sun, 11 Oct 2026 13:39:52 +0700 https://learning-notes-dz2.pages.dev/posts/2026-10-20/ Sat, 10 Oct 2026 08:44:00 +0700 https://learning-notes-dz2.pages.dev/posts/2026-10-20/ A fault in an agent should be corrected at the lowest layer that can still fix it, because that layer is the fastest to change, check and undo. Anything above that layer which rewrites itself is only safe if every change has to pass a gate that can say no. https://learning-notes-dz2.pages.dev/posts/2026-10-17/ Sat, 10 Oct 2026 08:43:00 +0700 https://learning-notes-dz2.pages.dev/posts/2026-10-17/ A judge that always picks position A is perfectly consistent, and an agent that abstains on everything never makes a confident mistake. Every evaluation number needs a degenerate baseline before it means anything. https://learning-notes-dz2.pages.dev/posts/2026-10-16/ Sat, 10 Oct 2026 08:42:00 +0700 https://learning-notes-dz2.pages.dev/posts/2026-10-16/ An agent is only as safe as the place where its authority sits and the strength of whatever checks its output. Filters, loop design and human approval all reduce to those two questions. https://learning-notes-dz2.pages.dev/posts/2026-10-13/ Sat, 10 Oct 2026 08:41:00 +0700 https://learning-notes-dz2.pages.dev/posts/2026-10-13/ A memory item that was written carelessly will mislead the agent, and a good item sitting behind a broken filtered search will never be found. Both failures return an answer instead of an error, so the only defence is a hard check at write time and another at read time. https://learning-notes-dz2.pages.dev/posts/2026-10-11/ Sat, 10 Oct 2026 08:40:00 +0700 https://learning-notes-dz2.pages.dev/posts/2026-10-11/ Appending to the end of a prompt costs almost nothing, and every rewrite above that point pays twice: once in cache price and once in whether the agent still knows what it was told. Layout and compaction are the same decision seen from two sides.