Latent.Space opened AI Engineer World's Fair with this and I haven't stopped thinking about it.

Six nested loops from token prediction to software factories.

Each level exits on a different condition, at a different timescale.

Most teams I work with are building at level 3 (agent turns: call tool, feed result). The jump to level 4 (goal loops: run, judge, retry) is where things get interesting. That's where the agent starts evaluating its own output.

Level 5 is where I live. Cron, subagents, multi-agent orchestration. Still figuring out level 6.