ICML 2024 paper RAFA reframes LLM agents as a closed-loop planning system
zhaoran_wang · x · 2026-07-29
RAFA proposes a closed-loop framework for autonomous LLM agents
The ICML 2024 paper “Reason for Future, Act for Now” (RAFA) frames LLM agents as a closed-loop learning + planning + acting system.
- The agent uses a memory buffer to learn from past feedback.
- It performs long-horizon planning for future trajectories, then acts now on the first step.
- After each interaction, it stores the new state/action/reward and re-plans from the updated memory.
- The paper claims sample-efficient interaction and provides theoretical regret guarantees.
The post argues that RAFA looks prescient in light of today’s interest in autonomous agents, online RL, test-time adaptation, and memory-based reasoning.
More from coding & agent
- Google pitches an end-to-end agent stack for build, test, deploy and govern — Saboo_Shubham_ · 2026-07-29
- Andrew Ng’s 12-page playbook maps four design patterns for multi-agent systems — theomitsa · 2026-07-29
- One GPU, many VMs: a practical guide to passthrough, vGPU, and API remoting — LoganGrasby · 2026-07-29
- Maven’s trending workshop teaches how to build AI agents from first principles — hugobowne · 2026-07-29
- Unreal Engine remains hard for agentic workflows because files, code and docs fight back — nptacek · 2026-07-29
- AI agents won’t replace you, but they can automate repetitive computer work — dotey · 2026-07-29