# RL experiment tracking for rollouts and rewards

A reward curve tells you something changed. It doesn't tell you which rollouts changed or why. Probe records the rewards per step and what happened inside each rollout, so you and your coding agent can trace a jump or a collapse back to the samples behind it.

Tell your coding agent:

```
Install Probe on this machine and its coding agents. Read https://research.prbe.ai/install.md and follow it.
```

## What you can follow

- **Rewards through training.** Log rewards and any other metric per step with `probe.log()`, next to the run's config and code.
- **Rollouts, turns and tool calls.** Record what happened inside a run as spans, then read the tree back with `probe span list` or the MCP `trajectory` view.
- **Harbor trials.** Capture Harbor trial directories as trials linked to their rollout spans, step by step.
- **Per-sample detail from Miles.** The Miles connector records sample-level measurements alongside the run.

```
probe span add $RUN --type rollout --name rollout-0 --step 1
probe trial add $RUN jobs/my-job/trials/swe-fix__x1 --step 600
```

See [Spans, trajectories and trials](https://docs.prbe.ai/tracking/trajectories) and [Framework integrations](https://docs.prbe.ai/sdk/integrations).

## Trainers it wires into

Ask your coding agent to run `/instrument-code` on your training script. The [instrument-code skill](https://github.com/prbe-ai/probe/blob/main/plugins/probe-daemon/agent-skills/instrument-code/SKILL.md) wires Probe into Miles, TRL, verl, SkyRL and ART trainers and Harbor jobs, so every rollout is recorded as a sample, including trainers that already have their own tracker.

## Why it helps

- **You can see why a run moved.** Open the rollouts at the step where the reward jumped.
- **Your agent learns from the last run.** It reads past trajectories before it changes the reward or the data.
- **Crashes don't cost you the record.** Data already saved stays available after the run or the machine stops.

## Learn more

- [Spans, trajectories and trials](https://docs.prbe.ai/tracking/trajectories)
- [An agent-native Weights & Biases alternative](https://www.useprobe.ai/wandb-alternative)
- [Run and track a swarm of research agents](https://www.useprobe.ai/research-agent-swarms)

Source: https://www.useprobe.ai/rl-experiment-tracking
