Limits and cost

Per-run budgets

Each agent limits its runs. The defaults:

Limit	Default	When it is reached
Max Model Calls	10	The run pauses with BUDGET_EXCEEDED.
Max Tool Calls	15	The run pauses with BUDGET_EXCEEDED.
Max Total Tokens	250,000	Checked before each model call. The run pauses with BUDGET_EXCEEDED.

They count the whole run. For an agent with Wait for Reply, they count each reply on its own.

A model call retried after a temporary error does not count toward Max Model Calls. Max Retries limits it. Its tokens still count. When the provider reports no usage, for example after a timeout, Core estimates the input at four characters per token and shows the attempt under Unmetered Model Attempts on the run.

Model cost

Set Input and Output Price per Million Tokens on the Agent Model. Each model call then records an estimated cost. Tokens per run are on the Agent Run record: Input Token Count, Output Token Count, and Total Token Count.

Salesforce cost

Each model turn is one Queueable job. A Stays in This Org tool call runs in the job of the model turn or Reads Outside Systems call before it. Every other tool call is one job.

Item	Cost
Start of a run with an identity, by someone else	2 jobs and 1 API call
Start of a run with an identity, by its execution user	1 job
Each model turn, with the Stays in This Org tool calls of its answer	1 job
Each Reads Outside Systems tool call, with the Stays in This Org calls after it	1 job
Each Changes Outside Systems tool call	1 job
Each Stays in This Org call after a Changes Outside Systems call or an approval in the same answer	1 job
Reply, approval, or rescue on a run with an identity	2 jobs and 1 API call
A typical run with 4 model turns and 3 Stays in This Org tool calls	About 5 jobs
The same run with 3 outside tool calls	About 8 jobs
Recovery, every 10 minutes	Up to 25 commands, and up to 10 extra passes when more are due
Retention, daily	Clears about 10,000 run payloads and deletes about 5,000 runs

How many runs an org can handle

The daily async Apex limit is 250,000 jobs, or 200 per user license, whichever is larger. About 30,000 runs a day with outside tools, or about 50,000 with Stays in This Org tools, use all of it. Other async work in the org shares the same limit.

When the org gets close, the Agent Core Health card shows Reduce load.

Large data volumes

Recovery, health, and retention filter on these fields: Status__c, RecoveryCheckAt__c, DispatchDueAt__c, PurgeAfter__c, and DeleteAfter__c. Above about a million Agent Command or Agent Event records, ask Salesforce Support for custom indexes on them.

Time

A model call takes from a second to two minutes, depending on the model and the answer’s length. Each job waits in the async queue before it starts, usually under a few seconds. The run story shows how long each step waited and ran. A tool call that ran in its model turn’s job waited 0 ms.