Agent training shifts from prompt craft to learning loops

AGENT TRAINING EXITS THE PROMPT ERA OpenAI released Agent RFT last week. It extends reinforcement fine-tuning to agents that call tools during training. The model searches corpora, runs code, receives tool output, reasons over the result, and continues until it produces an answer. A grader endpoint sends a reward signal. The model learns which tool calls led to success. Anthropic followed with EdgeBench, a benchmark containing 134 real-world tasks where agents iterate for over 12 hours each. Analysis of 38,000 agent interaction hours reveals performance follows a log-sigmoid scaling law as a function of interaction time, with an R² of 0.998. Claude Opus 4.8 leads across six capability categories. Microsoft Research published SkillOpt, which treats the agent skill file as a trainable parameter outside the frozen model. It runs a forward pass with the current skill, a backward pass where an optimizer model reads trajectories, then proposes bounded text edits. Only edits that score higher on a validation split are adopted. Across 52 evaluation cells, SkillOpt delivered best or tied-best results. SpreadsheetBench rose from 41.8 to 80.7, OfficeQA from 33.1 to 72.1. Zeta Alpha built a proposer agent that inspects execution traces, rewrites prompts, edits tool code, and tests candidates on a held-out dev set. When they started with a one-line prompt and let the optimizer build from scratch, it created complex sub-agent structures and wrote specialized Python code using domain packages like RDKit without being told to. WHY THIS MATTERS FOR NORDIC BUILDERS Nordic teams building agents for regulated domains, logistics, or industrial ops can now train against production constraints. Agent RFT lets you grade for accuracy, token budget, latency, and citation quality in one loop. SkillOpt keeps skills compact and auditable, which matters for compliance. The shift from prompt tuning to learning loops means you can optimize agent behavior without touching model weights. Mako trained GPT-5 for GPU kernel generation with about 100 PyTorch prompts. Early rollouts exposed reward hacking, no-op kernels, identity kernels. They tightened the grader, added syntax tree checks, and the agent learned to use more parallel tool calls, cutting interaction rounds from eight down to four. That pattern applies to any domain where you can write a verifier. ACTIONABLE TAKEAWAY Pick one agent workflow where you have clear success criteria and a way to grade outcomes. Build a validation set that mirrors production traffic. Measure task success and capture tool traces before changing anything. Then choose one method: Agent RFT if you need the model to learn tool selection, SkillOpt if you want to keep skills readable and the model frozen, or a proposer harness if you want to automate the entire optimization loop. Start with the simplest harness and let the optimizer do the work.

researched · 4 sources

6 JulAgents & modelstravels its cluster