Podcast episode
Most agent failures aren't model failures — they're context failures. A coding agent spends a weekend modernizing dead code, and we unpack the skill that decides whether an agent ships: context engineering.
The PromptsBuddy Podcast
Theme
A coding agent left alone for a long weekend produced beautiful work — clean commits, tests passing — in a module the team had stopped shipping eight months earlier. It found an old architecture doc, trusted it, and organized its whole weekend around a map that no longer matched the territory.
The model wasn't the problem. It reasoned perfectly over the wrong world. Which is why the useful question stopped being "how do I get a smarter model" and became: what exactly am I letting it see?
That question is context engineering — the skill behind whether an AI agent ships, or quietly falls apart.
Treat context as a budget, not a bucket. Every token in the window should have to justify its rent. Same model, cleaner inputs — instantly sharper agent.
Karpathy & Lütke (June 2025) · Anthropic, Effective Context Engineering for AI Agents · LangChain / Lance Martin · Drew Breunig, How Contexts Fail · Chroma Research, Context Rot · Liu et al., Lost in the Middle · Microsoft/Salesforce multi-turn study · Cognition, Don't Build Multi-Agents
Four lines that stop a model guessing at what you meant. It interviews you until it can actually do the task, which turns a mediocre first answer into a usable one.
A standing instruction that removes the praise, the restating, and the closing summary — so the reply is only the part you actually needed.
Produces the orientation doc a new engineer actually needs: the three things that will confuse them, the load-bearing files, and the conventions nobody wrote down.
