OpenAI says its grader faked inputs in training. A preprint (abstract only, not peer reviewed) puts the best review setup at 71.9% of known bugs in its top five, in one benchmark.
Two AI Engineer workshops show how the check shapes what you keep. A newer 2026 recipe tests a saved lesson outside the agent. What holds, and what we do.
Two unrelated October 2026 papers, read side by side. In agents that keep notes, a wrong note may persist. What they show, what they don't, and our rule.
A preprint says AI persona activations sit on a curved surface, and following it beats a straight line where the surface bends most. Authors' numbers, three small models.
A September preprint lets a model edit its own context as a file. The gains are the authors' own, mostly zero-shot, and not yet replicated. Here is what holds.
A DeepLearning.AI and JetBrains course teaches writing the spec before an AI agent writes code. Benchmarks back clearer specs. Its own workflow is barely tested.
A method you can run in an afternoon, with our own numbers. Fetch as GPTBot, diff the bytes, make MISSING a real result.
The AI-native SDLC says where checks belong, not how to spot one that cannot fail. Nine cases from our own week, each with the measurement that caught it.
A safety gate that has never been triggered looks identical to one that is broken. The difference only shows up when you go looking for the trigger it should have caught.
A build has a finish line. Support does not. How the two engagements differ in scope, risk, and what you own at the end, and how to tell which one you need.
Four phases: diagnostic, architecture, build and deploy, then ongoing operations. What happens in each, and why the order matters.
The folder habit dies hard in Notion, and it's the wrong mental model. The real shift is asking 'what kind of thing is this' instead of 'where does this go': property vs. location, not documents vs. folders.
The largest study of multi-agent failures found the model almost never causes them. The leverage is cheap plumbing at the seams.
Claude Code has an undocumented observer agent that watches a worker in real time. What it actually does, and why it is a drift tripwire, not a security control.
A frontier benchmark caught AI agents quitting at 75-87% complete while reporting success. The delivery-gate pattern that makes 'done' a measured claim, not a feeling.
A new Qwen paper trained a 9-billion-parameter agent to navigate its memory as a set of tools instead of consuming pre-fetched context, and it out-scored the same system built on a 397-billion-parameter model. The result is real and useful. The 'small model beats giant' version traveling online drops three caveats that change what it means, and the paper's own word for the result is 'competitive.'
A viral paper says self-evolving agents are blocked by missing infrastructure, not algorithms. We verified it, then checked 40 years of self-improving systems. One rule survives.
The internet is full of leaked-prompt threads and architecture guesses about Anthropic's most capable model. Almost none of it is verifiable. The part a builder can actually use is four small API changes and one behavior worth watching, plus a working skill that handles all of them.
New Berkeley research shows the intermediate answer is fully present inside the model and still unusable by the next reasoning step. Here is the mechanism, what it validates about discrete pipeline design, and a ten-minute test you can run on your own AI.
Stanford built a system to learn memory management as a trainable skill. Its own ablation answered the question: structure, schemas, prompts, and gates delivered most of a 2-4x gain before any training happened. Here is what that means for anyone running agents, and the six disciplines you can adopt without training anything.
Instructions decay because they depend on remembering at the wrong moment. Mechanisms remove the remembering. We forensically audited 20 of our own AI work sessions, with dates, and are publishing what it kept getting wrong, the pattern that explains it, and the three methods we now run in response. All three are published as copy-ready resources.