Agent "loops" in AI — the next step: continuous agents that improve code

Friends, a quick share from the AI world: at @Scale, Boris Cherni said agent "loops" are a real trend.

Key points:
- Loops: agents launch other agents and continuously improve code — PRs, refactoring, merging abstractions.
- It resembles recursive cycles, but termination is decided by a sub-agent — non-deterministically.
- The approach uses significant test-time compute: continuous runs burn tokens quickly and can be costly.
- Controls are needed: clear objectives, cost monitoring, drift prevention.

Why it matters: such autonomy can accelerate development but requires risk and cost governance.

Are teams ready to entrust critical systems to continuous loops?

#AI #agents #DevOps #technology