Quick answer
Amp meters agent work through credits and, for sustained runs, an hours-based measure tied to Orb. The hours reflect how long agent compute ran rather than how many messages you sent, which is why a single long autonomous task can cost far more than a dozen short prompts. The account panel shows the balance and the rate that applies.
What the hour measures
| Unit | Measures | Grows with |
|---|---|---|
| Messages | Prompts sent | How often you type |
| Tokens | Text processed | Context size and output |
| Orb hours | Agent compute time | Long autonomous runs |
| Credits | Spend | Any of the above |
The point of an hours measure is to price sustained autonomy, where the agent keeps working after you stop watching. A ten-minute run that reads files, edits, and tests is doing far more than a one-line question.
Estimate a run before starting
- Define the end condition for the task.
- Scope the files the agent may touch.
- Run a small version first and note the charge.
- Multiply for the full task, then decide whether to proceed.
- Stop the run when it loops instead of watching the meter.
Reduce hours without losing work
- Give the agent a precise task instead of an open-ended goal.
- Point at specific files rather than the repository root.
- Avoid re-running the same failing command.
- Break a large task into checkpoints so a bad run stops early.
- Use a cheaper mode for exploration.
UsageAtlas keeps Amp's reported balance beside local token history, so an expensive run is visible the same day. The Amp credits guide covers the balance side, and Amp agent credits explains workspace pools.
Questions, answered
What is an Orb hour in Amp?
A measure of sustained agent compute time. It grows with how long an agent runs, not with how many messages you send.
Are Orb hours the same as credits?
No. Credits are the spend unit; hours describe the work that generated the charge. The panel shows the balance in credits.
Why did one prompt cost so much?
A single prompt can start a long autonomous run. The cost follows the run length and context, not the prompt count.