An action limit should constrain the operations a voice session can actually execute. Asking the model to stop after a certain number of actions leaves concurrency, retries, and background jobs outside the control.
Action budgets need a defined unit
A session action budget caps a class of work, such as writes, external messages, or transfer attempts. Define whether a repeated request, a retry, and a rejected validation attempt consume that budget. Counting every tool invocation equally can punish harmless clarification while allowing an expensive compound operation.
For a proposed design, enforce limits in the execution service using trusted session and tenant identifiers. Reserve allowance atomically before starting an operation, then settle it according to the documented policy. Background jobs inherit the operation’s budget context rather than starting with a fresh allowance.
Use separate budgets where effects differ. Read-only lookups, record mutations, outbound notifications, and monetary actions carry different operational costs and consequences. A single global number is easy to implement but can be difficult to explain.
Exercise a fictional looping caller
A caller repeatedly asks to resend a confirmation message. Suppose the pilot policy permits two sends for the same request. This number is illustrative, not a universal recommendation. A third request should produce a clear explanation and an approved alternative, not another message.
Now run two simultaneous send attempts when one allowance remains. Only one should receive permission to execute. If both pass a local counter check, the limit is advisory rather than enforced.
| Scenario | Expected budget behaviour |
|---|---|
| Same operation retried safely | Follow the defined retry accounting rule |
| Two concurrent writes at the boundary | At most the remaining allowance executes |
| Session reconnects | Budget does not reset accidentally |
| Tool starts a background task | Task retains the original budget scope |
| Limit reached | Offer an approved non-writing fallback |
Inspect the destination system as well as the counter. A budget record showing two sends is wrong if three messages actually arrived.
Limits should support recovery
A hard cap can prevent runaway activity but also interrupt legitimate multi-step requests. Reserve enough allowance for required cleanup or compensation where the workflow needs it. Otherwise the limit may block the very action needed to repair a partial failure.
Expose limit events to operations without disclosing internal control details unnecessarily to callers. Repeated hits may indicate prompt loops, integration retries, abusive use, or a budget poorly matched to normal work.
This control is proposed. The voice platform case study provides related context, and partial-success handling explains why cleanup needs its own policy.
Session action FAQ
Should failed validations consume write allowance?
Define it explicitly. They may need a separate abuse limit even when no mutation occurs.
Can callers reset limits by reconnecting?
Avoid accidental resets by choosing the appropriate interaction, account, and time-window scopes.
What should happen at the cap?
Stop the governed effect, state the available next step, and preserve enough context for authorised staff to continue.