Dry run (for agent actions)

What it is

A rehearsal. The agent submits exactly what it intends to run, and the platform answers: are the arguments valid, is this key allowed to do it, is the connection there, does it need someone's confirmation, and how many billable calls would it make? Nothing reaches the vendor.

Why it matters

Agents guess. Anthropic's tool-use documentation notes that a model may fill in a plausible value for a missing parameter rather than ask. Writes in real systems are often irreversible. A dry run turns "try it and see" into "check it and know", for free, before anything happens. It also lets an agent show a person what it is about to do and what it will cost.

How it works

1. Validate the arguments against the action's schema.
2. Resolve the connection and confirm it exists for this tenant.
3. Check the key's permissions and cap headroom.
4. Report whether the action needs confirmation or an approval.
5. Estimate the upstream calls and the cost.

Example

Before creating forty tickets from a spreadsheet, an agent dry-runs one. The answer says the priority value is not one of the allowed options and that creating tickets on this key needs a person's approval. The agent fixes the value and asks the service-desk lead before running the batch.

Common pitfalls

• Dry runs that call the vendor. A check that reads from the vendor costs quota and money.
• Treating a pass as a promise. A dry run checks what the platform knows. The vendor can still refuse on its own rules, such as a closed accounting period.
• Checks that differ from the real run. If the dry run and the execution validate differently, the check proves nothing.

How actionlake handles it

The check tool in actionlake dry-runs any do or ask payload, including a whole plan. It reports every problem at once: missing or invalid inputs, missing connections, steps the key may not run, every approval the run would stop at and who would be asked, and the worst-case cost. For a tenant with a Dedicated IP it also says which address the calls would leave from, and whether that path is working. Nothing is sent, and it is free. Pair it with the free cost tool for spend so far and cap headroom. See the docs.