Long-running Agent isn't one thing

Several things have piled up in the last ten days.

Anthropic shipped a mode in Claude Code called /goal — "give it an objective and it runs until done." Cognition's Devin closed another round. China's Manus had its $2-3B Meta acquisition blocked by China's NDRC last month, and the deal is now in limbo. Karpathy gave a Sequoia keynote and named the whole thing "agentic engineering" — coining the term in the talk itself. Sequoia put out a piece titled 2026: This is AGI with a line I keep coming back to: "long-horizon agents are AGI; 2026 is their year."

Sitting with all of this, one thing stands out:

Everyone is using "long-running agent" and adjacent terms to describe wildly different things.

We did the same two years ago. AutoGPT. BabyAGI. 100K GitHub stars in a month. Headlines about "AI starting to move on its own." A year later it was all cold. Loops stuck, tokens torched, output hallucinated. Nobody remembers exactly how they died — only how loud the hype was.

If we don't disambiguate the words now, we're going to step on the same rake.

So this post is about pulling the words apart. Once the words are clean, we can talk about what's actually hard, how to pick tools, and which flavor I happen to be running.