What should count when an agent uses an operating environment?
Four ways to define a commercial unit between work requested and work finished. Each draws a different boundary around the evidence required.
Illustrative only. Active billing is not implemented. No prices, balances or production rates are set here. These examples do not select a final business model.
Resource cost and commercial unit
An environment can consume resources without resolving a problem. A resolved problem can still lack caller acceptance or evidence of a business outcome.
Underlying resource cost
Actual inference, execution and storage usage, supported by provider or executor receipts. Keep quantities by resource class and use real supplier or infrastructure cost records where available; missing cost evidence remains unknown.
Commercial billable unit
The agreed event a buyer would pay for: consumption, readiness, accepted resolution or an externally attested outcome. It is a contract definition, not a synonym for the resources consumed.
A usage-based agreement may map directly to measured resources. Other models still incur resource costs. No conversion from usage to price, revenue or margin is assumed.
Compare the evidence boundary
Each row defines one illustrative model. On smaller screens, focus the table region and scroll horizontally to read every column.
Billable events, measured units and required verification
A defined problem passes its agreed resolution checks and the named caller accepts the result.
One accepted resolution of a uniquely identified problem under a versioned acceptance contract.
Reproducible before/after checks and result provenance, followed by an explicit acceptance receipt from the authorized caller. Silence is not acceptance.
A separately agreed business outcome is externally attested under its attribution and observation-window rules.
One uniquely identified, attested business outcome; define eligibility, attribution and reversal terms before work starts.
An authorized external system-of-record receipt or independent attestation, with outcome identity, timing and agreed attribution evidence. An agent's self-report is insufficient.
The rules behind each unit
Retries and failure handling belong in the agreement before work begins. The policies below are examples, not implemented settlement behavior.
Usage-based
Suitable buyer / agent context
Platform teams and agent operators running variable workloads with explicit consumption limits.
Example
A data agent prepares a workspace, transforms a document and retains an artifact. Record actual inference, execution and storage separately, wherever consumption occurs.
Deduplication & retries
Deduplicate receipt IDs and overlapping intervals within the same tenant, resource and invocation or meter stream. Separate concurrent work and genuinely executed retries retain their own consumption; replaying a receipt adds nothing.
Failure handling
Verified consumption may still qualify when a task fails. Missing evidence stays pending. Any excluded provider faults or retry allowance must be agreed in advance.
Tradeoff
Direct visibility into consumption, but the buyer carries task-efficiency risk. A verified usage record does not prove a useful result.
Formation-based
Suitable buyer / agent context
Agent infrastructure buyers who need dependable operating conditions while their calling AI retains task ownership.
Example
A coding agent receives a prepared workspace and a checked execution capability. The environment is ready; no code fix or accepted result is implied.
Deduplication & retries
Use tenant, formation request and agreed scope revision as the unit key. Rechecks, reconnects and repair within that scope do not create another unit. A scope change needs fresh authorization.
Failure handling
Blocked or failed formation creates no readiness unit. If readiness is invalidated before the agreed handoff, hold or reverse that candidate; later repair follows the agreed service terms.
Tradeoff
Makes the environment boundary explicit, but readiness alone says nothing about resolution. Ongoing resource use needs its own inclusion or metering terms.
Resolution-based
Suitable buyer / agent context
Engineering or operations buyers delegating a bounded, testable problem to an agent.
Example
An agent repairs an invalid data export. Schema and reconciliation checks pass, then the calling system explicitly accepts the corrected artifact.
Deduplication & retries
Key by tenant, problem ID and acceptance-contract version. Multiple attempts produce at most one accepted resolution. Reopening within the agreed window is repair, not a second unit.
Failure handling
Unresolved, rejected or unaccepted results create no resolution unit. A material regression within the agreed window holds or reverses the candidate pending review.
Tradeoff
Connects the unit to a bounded useful result, but requires clear acceptance ownership and a reopening window. Resolution does not establish downstream business value.
Outcome-based
Suitable buyer / agent context
Business buyers and delegated agents whose result can be independently observed outside the operating environment.
Example
A reconciliation agent corrects an exception. A buyer-authorized finance system later attests the defined reconciled record within the agreed window; an internal 'done' message cannot qualify it.
Deduplication & retries
Deduplicate the external outcome ID per tenant and agreement. Retries, multiple agents and repeated attestations cannot count the same outcome twice; competing claims require review.
Failure handling
Absent, disputed or out-of-window attestation leaves the outcome unqualified. Reversed business events require a linked reversal under the agreement, never a silent rewrite.
Tradeoff
Aligns the unit with an observed business event, but introduces attribution disputes, delayed evidence and factors beyond the agent's control. Caller acceptance alone is insufficient.
Same work, different unit
Follow one illustrative lifecycle. Selecting a model changes only the explanation on this page.
Local display only · Not billed. No work is run or metered.
Environment ready
The candidate is the checked readiness handoff for this scope, even if the caller's later task fails.
Form the environment
Infer conditions from Outcome or Observe; prepare only authorized resources. Real consumption can begin here.
Environment ready
Example unit boundary · Not billed
Readiness checks pass for the agreed scope. The calling AI still owns its task.
Perform bounded work
The agent attempts the task. Inference, execution and retained storage can consume resources, including on failed attempts.
Problem resolved
Agreed resolution checks pass with evidence. Caller acceptance remains a separate event.
Caller accepts result
The authorized caller records explicit acceptance against the agreed criteria and result revision.
External outcome attested
An external authority supplies evidence of the defined business event and its attribution. This may happen later, or never.
These stages explain different evidence boundaries, not a mandatory workflow. Storage metering can continue after acceptance until retention ends. Nothing shown is a live operating record.
Optional hybrid: readiness + consumption
A buyer could agree to one formation unit at the checked readiness handoff, plus separately measured inference, execution and storage consumption.
Formation component
One unit for the authorized scope revision, subject to readiness evidence and the formation retry and failure rules above.
Resource component
Admit actual resource receipts using the usage rules above. Define which formation resources are included, and exclude those receipts from the additional usage component so the same consumption is not counted twice.
Before adoption
Agree resource boundaries, authorization limits, retry exclusions, invalidation and reversal terms. If formation fails, there is no readiness unit; whether verified consumption qualifies must already be explicit.
This hybrid is one design option, not a recommendation. Production rates, commercial terms and any billing implementation remain undecided.
Evidence before a charge
Any future model would need an authorized, versioned agreement: who may accept, what qualifies, how retries are identified, when evidence becomes stale, and how disputes and reversals are handled.
A commercial rule cannot expand an agent’s permissions. Missing or conflicting evidence must remain visible. Business outcomes require external attestation; neither an internal success flag nor the agent’s own account can establish them.