Two Kinds of AI Agents (And Why You Need Both)

Persistent context agents vs. stateless decision functions. When to build which, how to evaluate each, and why the industry is conflating two fundamentally different problems. Andrej Karpathy ended his 2025 Year in Review with this: "I don't think the industry has realized anywhere near 10% of their potential even at present capability." I think the reason is that we're treating "AI agent" as one category when it's actually two. And the evaluation strategies, architectures, and failure modes are completely different for each. The Two Agent Archetypes Type 1: The Persistent Context Agent This is an AI that lives alongside you. It has memory. It knows your codebase, your preferences, your ongoing projects, what happened yesterday. It accumulates context over time and uses that context to be more helpful. Examples: Claude Code running on your machine with access to your files, git history, and environment A Slack-integrated ops assistant that knows your team's systems, reads channels, tra