Training Frontier Knowledge Work Agents: A 397B Open Recipe with SkyRL
summary
Mercor post-trains knowledge‑work agents with SkyRL and Harbor’s Archipelago harness, scaling from 35B to a 397B model on 20 H200 nodes via Ray, and publishes the open recipe and weights. The pipeline enforces near‑zero harness error rates, implements token‑in‑token‑out tracking, overfits 32 tasks before hero runs, and reports pass@1 improvement from 16.11% to 27.29% on APEX‑Agents with a transparent per‑step cost and system tuning breakdown.
AI-written summary · about
chapters (15)
- 0:00Introduction and background
- 1:07Meror and Apex Agents overview
- 3:42Headline results and open sourcing
- 5:07Six-step RL playbook overview
- 6:11SkyRL and Harbor stack explained
- 11:07Step 1A: Environment robustness
- 13:10Step 1B: Harness optimization
- 15:18Step 1C: Token in token out
- 18:31RL systems tuning
- 21:01Overfitting on subset of tasks
- 22:27Small-scale RL runs
- 25:59Hero run on 397B model
- 27:04Generalization studies
- 29:09Open sourcing and future plans
- 30:14Q&A session
related talks
27mSkyRL: A Modular RL Library for LLM Agents
19mSelf-Training Agents: Hermes Agent, HF Traces, Skills, MCP & Finetuning
07h 47mAI Agents Conference 2026 - Day 5
19mTraining Agentic Reasoners
42mComputer-Use Agents: From SFT to Classic RL
31mLayla El Asri, Research Scientist, Maluuba, NYC 2017
49mAI Agents: Collaboration in Real-Life Applications
24mAI Agent Swarms Are Amazing