Once a Digital Employee owns long-running, recoverable work, a budget is no longer just a reporting number. It becomes an authority boundary that determines whether the runtime may admit the next unit of model work.
Central judgment
The safer interpretation of budget exhaustion is neither failed nor immediate termination. It is a reversible work-admission state: new generative requests stop entering the runtime, already accepted work state remains durable, narrowly scoped settlement may continue, and execution resumes only when an authorized policy owner changes the budget.
This conclusion is produced from the Q-20260808-01 Research Object. Production did not reopen the Signal Pool or perform new analysis from the Reading Result.
Source
The sole analytical input is Research Object — Pause-Preserving Budget Admission. The Reading Result is cited only as the provenance path already declared by that object.
Observation
The Research Object separates several states that are easy to collapse in an operator UI: budget policy, exact enforcement accounting, rounded operator-facing usage, admission of new work, settlement of already-admitted obligations, and the durable state of the long-lived job.
That distinction changes the meaning of a “hard budget.” It can be a hard gate on admitting new model requests without being proof of an exact final invoice ceiling, because work admitted before the gate can still settle afterward. Likewise, a budget pause preserves job identity and history. It is neither business completion nor execution failure.
Comparison
| Runtime semantic | New generative work | Durable work state | Narrow settlement | Authorized resume |
|---|---|---|---|---|
| Running | Allowed | Preserved | Allowed | Not required |
| BudgetPaused | Blocked | Preserved | Allowed | Required |
| Failed | Blocked | Implementation-dependent | Usually blocked | Recovery path |
| Completed | Blocked | Preserved | No new work expected | Not applicable |
This table is a Research Center synthesis from the Research Object; it does not claim that the source product exposes an identical named state machine.
Discussion
The important engineering mechanism is not budget display. It is the conversion of economic authority into a runtime admission decision. Budget policy should attach to the long-lived WorkOrder or equivalent job identity, and only an explicit policy owner should be able to change it. A worker retry, reconnect, or model-generated assertion must not silently reopen paused work.
Exact enforcement accounting should also remain distinct from rounded operator display. The Research Object explicitly preserves this boundary: a visible usage number may be approximate, while the admission gate must rely on the enforcement quantity. Promising an absolute final spend ceiling would exceed the current evidence.
Shared session budgets introduce another governance gap. If several threads share one cap, the mechanism does not automatically provide role-, department-, or task-level accountability. Those higher-level economic controls need their own policy objects.
Engineering impact
For Digital Employee runtimes, separate Active, BudgetPaused, WaitingForAuthority, and Completed. Persist policy owner, enforcement usage, display usage, pause reason, and resume authorization separately, and keep a narrow allowlist for settlement actions that close already-started obligations.
For CodeFlowMu, this should first appear in Runtime and WorkOrder state plus observation surfaces, not as a protocol expansion derived from one product mechanism. Recovery should reopen execution from accepted durable state rather than depend on the original provider session remaining alive.
Boundaries and counter-evidence
The evidence does not establish an exact final-cost ceiling, invoice alignment between list cost and negotiated spend, per-thread budget partitioning, exactly-once event delivery, reconnect guarantees, or organization-wide cumulative budgeting.
The claim here is therefore narrower: budget should govern work admission. A single budget field is not sufficient evidence for complete enterprise cost governance.
Future work
The next useful test is a minimal budget-state contract separating cap, exact consumed amount, displayed amount, settlement authority, and resume authority. It should also test composition of role budgets with a shared WorkOrder cap, crash recovery of BudgetPaused, settlement receipts, and operator decision UX.
Visualization note
The diagram models Active → Budget Gate → BudgetPaused → Settlement Drain → Authorized Resume. It is a Research Center synthesis based on the cited Research Object and does not imply unsupported quantitative results.
Evidence and references
- Research Object — Pause-Preserving Budget Admission: sole analytical input for Production, including judgment, uncertainty, counter-evidence, and engineering impact.
- Reading Result — Session Budget Governance: provenance path declared by the Research Object; Production did not perform new research from it.