Handing your prices to automation takes trust, and trust built on "just believe us" is worthless. The activity log is GPU Autopilot's answer: every single pricing decision is recorded with the market context that produced it, so you can audit what the agent did, when, and why, down to the individual price write.
Why an audit trail matters for automated pricing
Most pricing scripts hosts run are black boxes even to their authors. The price changed overnight; was that the script, the platform, or a typo? When revenue dips you cannot tell whether the logic misfired or the market moved, so you either turn the automation off or stop looking. Neither is a way to run a business on Vast.ai.
GPU Autopilot's pricing loop is deterministic logic, no LLM involved, which means every decision has a specific, explainable cause. The activity log surfaces that cause. If you ever disagree with a decision, you can see exactly which input drove it and adjust the setting responsible, instead of guessing.
The eight action types
Every entry carries a color-coded action type. Eight cover everything the agent and your machines do:
| Action | Color | Meaning |
|---|---|---|
| Updated | Blue | The agent changed your listing price |
| Held | Gray | The agent evaluated the market and kept your price where it was |
| Skipped | Yellow | The agent did not act on this machine this cycle |
| Failed | Red | An action was attempted and did not go through |
| Rented | Green | A renter took the machine |
| Available | Sky blue | The machine became available again |
| Schedule On | Purple | A day-of-week floor override engaged |
| Schedule Off | Light purple | The machine returned to its normal floor |
The colors make the feed scannable at a glance: a wall of gray and blue is a healthy machine in a moving market, green and sky blue tell the rental story, and any red deserves a click.
Reading an Updated entry: a worked example
Each entry includes the market snapshot at the moment of the decision: your price, the competitor price, the median, and the clearing price. Here is how to read a typical Updated entry on an RTX 4090 listing:
- Your price: $0.38. Where the listing stood before this cycle.
- Clearing price (rented median): $0.35. The price where half of the active rentals for comparable machines are trading. This is the agent's primary target, built from what renters actually pay rather than unsold asking prices.
- Cheapest competitor: $0.34. A comparable machine listed below you and below the clearing price, which is the trigger condition for an undercut.
- New price: $0.32. The competitor's price minus your configured bid offset of $0.02, putting you first in the price-sorted results.
Every number in that story is visible in the entry. The next cycle, a few minutes later, might log a Held: same market, no competitor movement, nothing to do. And if you later become the cheapest by a wide margin, you will see an Updated entry moving your price back up toward the next competitor, because the offset works in both directions. If a sequence like this ever surprises you, the guide on strategy modes explains how each mode shapes these decisions.
What Held, Skipped, and Failed are telling you
The non-events carry as much information as the price changes:
- Held is the normal state. The agent runs every few minutes, but markets do not move that often. Long runs of Held mean your price is already positioned correctly and the agent is deliberately not churning it.
- Skipped means a cycle passed without action on that machine. Occasional Skipped entries are routine; a machine showing nothing but Skipped is worth a look at its settings.
- Failed means an attempted action did not go through. One Failed followed by a successful Updated is a hiccup. Repeated Failed entries point at something real, often an API key problem, and are exactly the kind of event worth routing to your phone.
Pair the log with alerts
Filtering, exporting, and monthly reviews
The feed filters by machine, action type, and date range, and the data is export-ready for your own analytics. That turns the log from a debugging tool into a review habit. A useful monthly routine: filter one machine to Rented and Available entries to see its true utilization pattern, then to Updated entries to see how often the market actually moved. Machines whose rentals cluster on certain weekdays are candidates for day-of-week floor overrides, and the Schedule On and Schedule Off entries will then let you verify those overrides fire on time. Ten minutes of reading the log usually teaches you more about your market than an hour of watching live prices.