I had local AI models play Zork, the 1981 text adventure, to study why agents struggle to navigate connected systems. One started responding in Thai. Most scored zero. All got hopelessly stuck in the maze. What broke says a lot about why agents get lost in microservices too.
AI tools are trained on extreme goal fulfillment: you ask "build X" and they build X. I've learned that prompting against that grain gets even better results: asking the model to stress-test ideas, not just execute them. It takes a deliberate permission structure to get there.
Workday's CEO called AI agent startups "parasites." Linear shipped a native AI agent that understands software development workflows. These are two ends of a spectrum, with most of the industry stuck in the middle — exposing data through MCP without embedding any intelligence. The companies that win will be the ones that make AI native to their domain workflow.
I learned algorithms from textbooks and papers, building mental models from pseudocode and hand-drawn sketches. The hardest part was never reading the algorithm: it was seeing what it was actually doing to the data structure in memory. This post is the tool I wish I had back then.
AI agents can generate code 100x faster, but for companies stuck in the "Unhappy Middle" — with legacy debt, bespoke frameworks, and zero slack — the bottleneck has shifted from writing code to verifying it. Here's how engineering leaders can cross the chasm by becoming gardeners, not janitors.
LLMs have statistical momentum: even when they know a new standard like Streamable HTTP exists, they often revert to the legacy patterns they saw most in training. Here is how to use "strong anchors" and "zero-prompt pruning" to keep your agentic systems from being haunted by 2024.