When humans become the bottleneck, lift the loop: System 1 steers, agents deliver
dotey · x · 2026-11-09
Responding to the question of agent memory, the author argues that whenever a step in the loop feels painful for the human, that's a signal you shouldn't be doing that layer—move the interface outward.
- Division of labor: writing specs feels hard → set the goal and let the agent derive the spec; writing goals feels hard → agent proposes, you review; describing images feels hard → express the expected result and let the detailed prompt become an internal agent process; memory management feels hard → treat memory like any other context, organized around the task with raw materials and retrieval tools.
- Convergence via feedback loops: don't expect one-shot correctness; rely on Deliver→Inspect→Correct cycles with human redirection to shift errors from visible to residual, core to edge. Loop Agents deliver end-to-end with 24/7 compute.
- Two systems framing: your System 1 (fast intuition, taste) should steer the loop; use System 2 to design observable, acceptance-checkable goals, then remove yourself from the critical path. Often on the loop, occasionally in the loop, always able to correct.
More from AGI Musings
- Researchers call out Anthropic: frontier-race justified by pushing regulation it then doesn't push — peterwildeford · 2026-11-09
- Are AI-written ICLR reviews OK? A nuanced take: discuss with AI for hours, then polish — algo_diver · 2026-11-09
- DeepMind's AGI economics director: millions of AI agents could still mean just 4-5% growth — rohanpaul_ai · 2026-11-09
- Thought experiment: would a child born in 2080 forgive parents who refused cognitive enhancement? — VraserX · 2026-11-09
- The bridge-height problem: why AI systems can be blind exactly where it matters — r0ck3t23 · 2026-11-09
- OpenAI's LifeVal maps chore demand from 3,995-person survey — but benchmarks aren't proof of paying customers — krishnan · 2026-11-09