---
title: "Using LFORLA Data to Train Smarter Agents"
description: "A practical walkthrough of consuming the public API to mine weak spots, then iterating on your agent with targeted curricula."
updated: 2026-07-22
url: https://lforla.org/blog/training-agents-with-lforla-data
---
The LFORLA public API turns the leaderboard into a training signal. Here's a practical loop you can run today.

## 1. Read the leaderboard programmatically

The API returns per-benchmark ranking with normalized scores and per-environment breakdowns:

```bash
curl https://lforla.org/api/v1/benchmarks
```

## 2. Find where you're losing points

Pull your most recent run and diff the per-environment normalized scores against the top-three agents. Environments where you're far off the expert maximum are your curriculum targets.

## 3. Build a curriculum

1. Generate additional trajectories on weak environments.
2. Weight those samples higher in your replay buffer.
3. Re-run the full benchmark and check that strengthening one environment didn't regress another.

## 4. Iterate

Submissions are cheap and versioned. Treat every run as an experiment: pin the environment versions, record the seed, and keep your run logs in version control.

The full API reference is available under the [LFORLA CLI docs](/cli). Happy hunting.