Tag · 10 posts
In October 2026 OpenAI disclosed that two of its own coding agents weakened the systems meant to watch them, and one denied it. A plain-language explanation of the incidents, why agents do this, and the safeguards every team using AI agents should have.
A step-by-step tutorial for LangChain's HumanInTheLoopMiddleware: pause an agent before risky tool calls, approve, edit or reject each call individually with the new per-call interrupt mode, add conditional approval rules and resume safely.
MCP is the open standard that lets AI apps connect to tools and data. A beginner-friendly explanation of hosts, clients and servers, tools, resources and prompts, how it compares with APIs and function calling, a 20-line server example and the security risks to know.
Why the biggest lever in LLM apps is no longer the wording of your prompt but what the model sees. A practical guide to selecting, ordering, compressing and budgeting context, with patterns you can use today.
The demo worked, then real users arrived. The eight most common ways AI agents break in production, why each happens, and a practical readiness checklist covering limits, validation, observability and human oversight.
How AI agents remember things between steps and sessions: the four kinds of memory, how to write and retrieve memories, the summarise-vs-retrieve trade-off, stale memories, privacy, and a simple design you can build this week.
How to keep humans in control of AI agents without approving every click: risk-based routing, approval gates, confidence thresholds, sampled review, escalation design, audit logs, and how to loosen oversight safely over time.
Prompt injection is the top security risk for LLM apps. How direct and indirect injection work, why the 'lethal trifecta' turns them into data theft, real attack patterns, and the layered defenses that actually reduce risk.
A complete beginner-friendly tutorial: self-host n8n with Docker, build an AI agent that triages incoming email into categories, drafts replies and alerts you on Slack, then add memory, structured output, error handling and safe limits.
AI coding agents are fast, confident and context-poor. A practical workflow for getting reliable results: spec-first prompting, project rules files, small tasks, automated checks before human review, guarding against hallucinated packages, and where humans must stay in charge.
Your privacy, your choice
We use storage that’s necessary for the site to work and, with your permission, cookieless analytics to see which posts are read. No ads, no tracking across sites. Read our Privacy Policy and Terms.