🗞️ CAIS and Scale (AI safety research group) say Fable 5 now automates 16.1% of real remote-work projects, about 2x Opus 4.8.

2025-08-21 · Source: Rohan's Bytes · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Intermediate, medium

Summary

A Hugging Face Engineer released a comprehensive slide presentation detailing how to train an open coding agent using TRL and OpenEnv. The core premise emphasizes training agents within the actual tool loops they will utilize, rather than solely for chat-based interactions. The presentation outlines a two-part agent setup, combining a model with a harness for tool interaction. Key training steps include Supervised Fine-Tuning (SFT) using agent traces and completion masking, followed by Group Relative Policy Optimization (GRPO) for reinforcement learning, where passing tests serve as concrete rewards. The methodology also integrates OpenEnv to provide agents with real-world environments for iterative learning and error correction, highlighting that agent training is partly a systems problem requiring async training with AsyncGRPOTrainer for long rollouts.

Key takeaway

For AI Scientists and ML Engineers developing coding agents, prioritize training within actual tool-driven environments rather than relying solely on chat-based instruction. Your development strategy should integrate Supervised Fine-Tuning with reinforcement learning via Group Relative Policy Optimization, using concrete test outcomes as rewards. This approach, coupled with real-world environment simulation like OpenEnv, will significantly improve agent robustness and practical utility, moving beyond theoretical chat performance.

Key insights

Coding agents require training within their operational tool environments, not just for conversational proficiency.

Principles

• Agent training needs real tool interaction.
• Reward successful actions with concrete outcomes.
• Asynchronous training supports long agent rollouts.

Method

Train coding agents via Supervised Fine-Tuning on agent traces with completion masking, then apply Group Relative Policy Optimization using test pass/fail as rewards within OpenEnv's real environments, leveraging AsyncGRPOTrainer.

In practice

• Implement SFT on agent interaction logs.
• Use test results for RL rewards.
• Integrate OpenEnv for realistic agent testing.

Topics

• Coding Agents
• Reinforcement Learning
• Supervised Fine-Tuning
• TRL
• OpenEnv
• AI Training Environments

Best for: Research Scientist, AI Engineer, Machine Learning Engineer, AI Scientist, Director of AI/ML, Consultant

Related on AIssential

• Data News — Week 25.43 — Modern AI systems combine pre-training and reinforcement learning (RL) to achieve advanced reasoning and behavioral capabilities.
• Harness engineering for coding agent users — Build trust in coding agents by combining feedforward guides and feedback sensors to anticipate and correct issues.
• Feedback-Driven Tool-Use Improvements in Large Language Models via Automated Build Environments — Automated environment generation and verifiable rewards significantly boost LLM tool-use performance via reinforcement learning.
• RL for Agents Workshop - Deep Dive on Training Agents with RL and Open Source — AI agents are evolving rapidly, enabling long-horizon tasks through advanced RL training, custom models, and robust evaluation.
• Training Agents: Live tutorial on how to fine-tune a coding agent for continual learning — The tutorial outlines a supervised fine-tuning (SFT) baseline for coding agents using public traces and TRL + LoRA.
• The Best Open 32B Model (Open Thoughts Agents) — Optimizing supervised fine-tuning data through a structured pipeline significantly enhances open-source AI agent performance and generalization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Rohan's Bytes.