A 5-step curriculum for post-training agents: SFT to RL in a real harness, code included
SergioPaniego · x · 2026-10-28
Sergio Paniego lays out a 5-step curriculum for post-training agents, from SFT to RL inside a real harness, with livestreams, code and docs for each step. The diagrams cover four training signals: SFT on agent traces, on-policy distillation scored per token by a teacher, group-relative RL (GRPO-style), and multi-turn RL environments scored by outcome. Rollouts run in remote sandboxes with generation and training on separate GPUs, updated via an async GRPO trainer.
More from coding & agent
- Developer shares his split: he designs the algorithm, DeepSeek makes it fast — cephaloform · 2026-10-29
- Microsoft execs detail how accelerators, codegen and custom agents transform internal dev — kevin_scott · 2026-10-29
- x401 Protocol Aims to Give the Agentic Web Verifiable Identity Proof at HTTP Level — csuwildcat · 2026-10-29
- ExtraHand Debuts as an Outer Loop to Route Multiple Models Into Your Coding Agents — cneuralnetwork · 2026-10-29
- Mirendil Launches Lor, Agents That Run Hundreds of Parallel AI R&D Experiments — shayan_ · 2026-10-29
- Oqoqo: hillclimb your MCP server with evals across 11 coding agents — pegasus013 · 2026-10-29