Research

Train an agent that writes RL training jobs for small models

A developer trained an agent whose job is to write reinforcement-learning jobs for small models, for about $1.3k, and open-sourced it.

By Danau5tin on Hacker News · Posted · Added Oct 16, 2026 · 2 sources

Difficulty
AdvancedOur rating
Time to build
Not reportedOur estimate
Monthly cost
Not reported
Revenue
Not reportedWe only show revenue the source states
Engagement
107 HN points49 comments · fetched Oct 15, 2026

What they did

The outer agent gets a task spec and writes a full training job; each job trains a small Qwen model on rented GPUs, and the improvement becomes the outer agent’s reward.

How it went

Open-sourced as ai-trains-ai; the title puts the cost at about $1.3k.

Worth knowing

Hidden evals stop the agent from gaming its reward.

Builder recipe

Our own starting point, written from the source. Adapt it to your setup.

Stack

  • Qwen
  • prime-rl
  • Runpod

Steps

  1. Define the question and success criteria.
  2. Let agents search, run or summarize in parallel.
  3. Verify key results yourself.
  4. Write up findings with sources.

Starter prompt

Tags

Industry AI

Read the original

Related use cases

The Daily 5.

Five real AI agent use cases each morning, each with a recipe and a source. Plus a Sunday roundup. Free.