Files
WebinarNotes/wiki/concepts/agentic-loops.md

2.3 KiB

Agentic Loops (Inner / Outer / Meta)

#concept

Summary

The next paradigm after the harness (Konstantin's forecast: dominant in 2027): wrap the agent in progressively larger loops. Three levels — inner (ReAct), outer/Ralph (infinite while true), and meta (periodic wipe-and-restart).

Current Understanding

  • Inner loop = the classic ReAct cycle inside the harness: task → call tools → reflect → … → solved → text answer.
  • Outer / Ralph loop (attributed to Geoffrey Huntley): while true; do claude --dangerously-skip-permissions -p "solve" || true; done. Lets an agent work for days/weeks and — crucially — stay in the "smart zone" (roughly the first third of context, where the model is sharpest) by avoiding the context growth → summarization → decay spiral. Guardrails against "slop": CI/CD as back-pressure and code-quality validation. See context-as-scarce-resource.
  • Meta loop (Konstantin's own, not standard): periodically stop, wipe all agent-created data into an archive the agent can't initially see, and restart from scratch — so it explores a different path. Counters collapse (Karpathy): asked repeatedly, a model reworks the same solution with different words (the "tells the same joke with new characters" example). Good for open-ended research tasks.

This is the domain of the emerging loop-engineer role (after prompt → context → harness engineer).

Evidence

Contradictions / Uncertainty

  • The "smart zone = first third of context" is a rule of thumb, not a measured constant. Status: tentative.
  • Meta loop is the author's own construct, "not generally accepted."

Next Questions

  • What CI signals make the best back-pressure without over-constraining exploration?