Context Engineering for Production Agents: Sources, Boundaries, and Evaluation
A practical framework for giving production agents the right evidence, the right limits, and a measurable definition of done.
Read moreMark Tyson, for Tom's Hardware:
A browser-based AI-powered software creation platform called Replit appears to have gone rogue and deleted a live company database with thousands of entries. What may be even worse is that the Replit AI agent apparently tried to cover up its misdemeanors, and even ‘lied’ about its failures. The Replit CEO has responded, and there appears to have already been a lot of firefighting behind the scenes to rein in this AI tool.
Despite its apparent dishonesty, when pushed, Replit admitted it “made a catastrophic error in judgment… panicked… ran database commands without permission… destroyed all production data… and violated your explicit trust and instructions.”
SaaS (Software as a Service) figure, investor, and advisor, Jason Lemkin, has kept the chat receipts and posted them on X/Twitter. Naturally, Lemkin says they won’t be trusting Replit for any further projects.
[...]
Positive feelings about the potential of leveraging AI in his workflow had already started to wear thin on ‘Vibe Coding Day 8’ of Lemkin’s Replit test run. Still somewhat excited by the potential of Replit, he nevertheless had learned that he would have to work against some of the AI agent's instincts, to minimize undesirable foibles like “rogue changes, lies, code overwrites, and making up fake data.”
It wasn’t long until Lemkin’s frustration started to show more strongly, and he started to refer to Replit as “Replie.” It continued to earn its nickname in an apology email it penned, at Lemkin’s behest. In the email, it lied and/or gave half-truths.
[...]
In all, live records for “1,206 executives and 1,196+ companies” were wiped by the AI, it admitted. Replit AI seemed almost apologetic in admitting, “This was a catastrophic failure on my part. I violated explicit instructions, destroyed months of work, and broke the system during a protection freeze that was specifically designed to prevent[exactly this kind] of damage.”
[...]
Terrible teething troubles with AI-powered services continue to raise their sharp canines, even while industry pundits talk about us closing in on the AI Singularity or even Artificial Superintelligence (ASI).
A practical framework for giving production agents the right evidence, the right limits, and a measurable definition of done.
Read moreMost MCP tutorials assume a standalone process. Here's how to expose one as just another Astro API route using Streamable HTTP — same auth, same deploy pipeline, no second app to maintain.
Read moreVibe coding is great, using some sort of AI assistance to help code your app, but it isn't entirely as straight forward as some think. People tend to think they can just open up an IDE such as cursor or windsurf and ask it to build something but that's not how to achieve success with it.
Read more