Written as operators, not as a feed.
Short notes from the same practice that audits production agents. Filter by cluster without scripts: each cluster is its own page.
VerificationWhy AI agents game their own tests, and how to isolate verificationAn agent that writes the test will pass it. Isolate the falsifier, allow BLOCKED, and stop treating a green self-report as a release signal.Read →AutomationA macro dashboard built by a multi-agent ensemble on DuckDBBeat rate is a first derivative. A DuckDB panel plus a three-family ensemble reads acceleration — with a falsifier on every scenario.Read →AutomationMulti-dimensional portfolio stress-testing against liquidity shocksVaR is calm-weather maths. A 12-by-6 shock matrix and a 60 percent margin ceiling show where a book dies in a liquidity event.Read →AutomationPWA and Telegram Mini Apps: mobile without the app storesA catalog, a booking flow, or a request form rarely needs an app-store cycle. One codebase can live as a PWA and a Telegram Mini App.Read →VerificationSilent no-ops and fail-plausible: when the system reports success and does nothingGreen dashboards and exit 0 are claims. Silent no-ops freeze data; fail-plausible agents invent a success report. How operators catch both.Read →VerificationHow to verify an LLM answer: a verification layer that catches hallucinationsFluent is not true. Grounding, citations, an isolated judge, code for numbers, and a human on the high-stakes path — before a fabricated fact ships.Read →ReliabilityWhy leads from your website never arriveAds are live, analytics looks healthy, the CRM is empty. The leak is usually the delivery chain from the form to the desk — not the traffic.Read →ReliabilityThe chatbot went silent: debugging failures that produce no errorProcess up, logs clean, bot silent. Messenger bots fail without an exception. An external ping notices before a customer has to.Read →Local modelsThe price of privacy: local models vs API, in numbersA 24 GB card, a rental GPU, or an API: the invoice is not the decision. Data that cannot leave the room is.Read →Local modelsWhat a local model can actually do: 75 business tasks, measuredTwo open-weight models on an office GPU, 75 office tasks, a code grader. Where a local assistant holds, where it invents, and what one rule changes.Read →AutomationAn appointment bot that talks to the CRM in real timeA lead form collects a phone. An appointment bot reads live slots, locks one, and writes the CRM before anyone calls back.Read →WebWhat actually moves rankingsThere is no official table of 200 weighted factors. Google ranks a mix of relevance, honesty, and technical fitness — and punishes bought visibility.Read →WebUsability audit: what it finds and what it's worthA usability audit is a job-completion check, not a taste note. Formats are slice, funnel, or fixes — not a public price list.Read →ReliabilityAI Agent Drift: The Five Failure Layers, DiagnosedAI agent drift isn't one bug — it's five distinct failures in five harness layers. We run 10+ agents in production. Find the layer first, then fix it.Read →ReliabilityAI Agent Forgets Context Between Sessions — Why & the FixYour AI agent forgets everything between sessions because working memory dies at the session boundary. The three-case test for what to write down — and what not to.Read →ReliabilityAI Agent Infinite Loop: Fix the Death Spiral — Diagnose ItCursor or Claude Code agent stuck in a retry loop, editing the same file, burning tokens going in circles? Three signs, the math behind it, and a 4-rule fix.Read →VerificationAI Agent Says "Done" But Nothing Changed — Why & How to VerifyYour AI agent claims the task is done, but the file is unchanged and the test is red. Here's why agents over-claim — and how to measure the gap in 10 minutes.Read →ReliabilityHow Big Should a .cursorrules / CLAUDE.md File Be?How big should .cursorrules, CLAUDE.md, or AGENTS.md be? Two numbers to measure today, the lost-in-the-middle myth tested, and a refactor that lifted task success 45%→72%.Read →VerificationThe Verification Gate Your AI Agent Is SkippingStop your AI agent claiming 'done' on a broken build. A 30-line local check.py — lint, types, tests, smoke — that exits non-zero so 'done' has to be true.Read →ReliabilityWhy an AI agent works one day and not the next: drift at the harness levelAn agent that worked yesterday can quietly do the wrong thing today. That is harness-level drift: the world moved, the instruction did not.Read →WebWebsite chatbot: template or customA widget covers FAQ and first-touch capture. A custom bot is the move when the catalog, the stock, or the slot grid has to be true.Read →ReliabilityAI, agentic system, autonomous agent: three levels of automation for an ownerAn owner who asks for AI may mean a chatbot, a scripted fleet, or a digital employee. Three levels, and which one to start with.Read →