A Multi-Agent Budget Needs Cost Attribution Beyond a Limit

Imagine a report that says, “This run used 2.3 million tokens.” That is an observation.

It is not yet a management decision.

The number does not say whether the work answered a high-value question, duplicated an existing subtask, spent a healthy reserve on verification, recovered from a tool failure, or continued after the task should have stopped. It also does not reveal whether a cheap result was safe, correct, or useful.

That is why a multi-agent system needs more than a limit. It needs cost attribution.

A budget and an attribution record answer different questions

A budget answers:

What may this task spend before it must narrow scope, pause, ask, or stop?

An attribution record answers:

What did the task spend, why did it spend it, and what did that spending produce?

Both are necessary.

A global limit can prevent runaway work, but it can also produce misleading pressure. A team may cut the verification step because it looks expensive, while continuing redundant research because the total is all it can see. Or it may increase every cap at once because a task is “important,” without distinguishing a cost increase from an expansion of time, tool access, or authority.

Attribution makes the tradeoff inspectable.

The telemetry already has useful building blocks

The OpenAI Agents SDK usage reference shows that an agent runtime can preserve both aggregate request and token counts and per-request usage entries. It also describes task- and turn-oriented usage data for tracing. Those are useful building blocks, not a complete governance model.

Similarly, the OpenTelemetry GenAI registry records observability signals around model requests, responses, tools, and usage, while directing moved definitions to the dedicated GenAI conventions repository. Attribute names and stability statuses can evolve. The durable design lesson is to observe resource use with selected usage metadata, while excluding prompts, tool arguments, and private task data from a cost report; telemetry conventions alone do not enforce that privacy boundary.

Measurement alone does not reveal value. A record must connect the resource signal to the unit of work and the decision it supported.

The Agent Cost Attribution Record

The Agent Cost Attribution Record is a proposed minimal ledger entry for a meaningful unit of agent work. It is not a pricing calculator, a hidden-reasoning log, or a surveillance record for individual contributors. It is a way to connect consumption to a bounded task outcome.

• Record element: Task and parent-work reference — Question it answers: Which user objective and work unit incurred this usage? — Why it matters: Prevents a large aggregate from hiding which branch consumed the capacity
• Record element: Consumption category — Question it answers: Was this exploration, execution, verification, recovery, coordination, or rework? — Why it matters: Separates useful diligence from avoidable repetition
• Record element: Minimal usage signal — Question it answers: What requests, time, tool capacity, or cost estimate are relevant? — Why it matters: Supports accounting without collecting private task content
• Record element: Decision purpose — Question it answers: What question, acceptance check, or uncertainty was this work intended to resolve? — Why it matters: Connects spend to user value rather than a model step count
• Record element: Dependency and duplication posture — Question it answers: Did the work depend on another branch, or repeat an existing result? — Why it matters: Makes coordination cost and redundant exploration visible
• Record element: Outcome — Question it answers: What artifact, evidence, decision, partial result, or blocker resulted? — Why it matters: Distinguishes consumption from accepted progress
• Record element: Next governance action — Question it answers: Should the system continue, reuse, narrow, re-plan, hold, or ask for more budget? — Why it matters: Turns measurement into a bounded operational choice
• Record element: Retention and sensitivity — Question it answers: What minimal reference is safe to keep? — Why it matters: Protects privacy and avoids retaining prompts or raw tool payloads by default

The record should attach to work, not to a person. It should document value-relevant outcomes, not internal chain of thought.

Classify why capacity was spent

Not all usage is equally valuable, and not all expensive work is waste.

• Category: Exploration — Healthy example: Independent research paths for a broad, high-value question — Warning sign: Several workers search the same ground because tasks were not differentiated
• Category: Execution — Healthy example: A bounded task uses the right model and tools for the required output — Warning sign: Extra capacity is consumed after the task has already satisfied its acceptance conditions
• Category: Verification — Healthy example: A result is checked before a consequential claim or effect — Warning sign: Verification is skipped because it makes the total cost look worse
• Category: Recovery — Healthy example: A known failure is diagnosed and a bounded alternative is tested — Warning sign: Blind retries continue despite an ambiguous or permanent failure
• Category: Coordination — Healthy example: Workers exchange concise, useful artifacts and decisions — Warning sign: Large outputs are repeatedly copied through multiple contexts
• Category: Rework — Healthy example: A material correction or new evidence requires an affected branch to be redone — Warning sign: A missing dependency record forces the system to redo unrelated work

The classification is not an automatic judgment. A team still needs task context and human review. But it transforms “the bill is high” into a specific question about system behavior.

Multi-agent capacity should be tied to task value

The CostBench preprint studies cost-aware tool planning in a travel-planning environment where paths can have different costs and conditions can change. Its domain and results are limited to that benchmark. Its framing is useful: agents must sometimes adapt not only to whether a path works, but to its resource consequences.

An Anthropic engineering account makes a complementary practical point: multi-agent systems can spend much more capacity than ordinary chat and need tasks whose value justifies that capacity. It also notes that coordination-heavy tasks are not automatically a good fit for parallel workers.

The implication is not “minimize tokens at all costs.” It is:

1. define why extra work is justified;
2. link the work to an acceptance condition or uncertainty;
3. retain enough capacity for verification and recovery; and
4. learn which task shapes produce accepted outcomes rather than volume.

Avoid the cost-control traps

Three traps appear often.

Trap 1: The cheapest path is treated as the best path

A low-cost response can still be unsupported, unsafe, or incomplete. Cost attribution should never become a reason to weaken verification, bypass required approval, or conceal uncertainty.

Trap 2: All expensive work is called waste

Some work is essential because it validates a change, checks a tool outcome, resolves conflicting evidence, or recovers from a legitimate failure. A healthy system can identify this as verification or recovery rather than penalize it as unexplained overhead.

Trap 3: Cost collection becomes data collection

A useful ledger does not need prompt text, private arguments, personal data, or hidden reasoning. It can retain task references, work categories, aggregated usage, and outcome summaries. If a detail is sensitive, record that an allowed validation occurred without copying the underlying material into a broad report.

Metrics that improve operations

Teams can track a small set of signals without pretending that one number measures intelligence:

• cost per accepted outcome: aggregate usage associated with work that met stated acceptance conditions;
• verification share: the portion of a task's capacity spent checking important claims or effects;
• duplicate-work rate: the share of work units that repeated an already sufficient result;
• recovery cost visibility: whether a failure and the bounded work spent resolving it are linked;
• unattributed-capacity rate: usage that cannot be connected to a task, work unit, or decision purpose;
• rework locality: whether a correction causes only dependent work to be redone; and
• budget-transition quality: whether a task narrows, holds, or asks at a defined boundary instead of silently overspending.

These measures are most useful when reviewed alongside quality, safety, and completion evidence. A lower cost is only a win when the outcome remains acceptable.

A practical adoption path

1. Choose one multi-agent workflow with a clear objective and an observable acceptance condition.
2. Give each meaningful worker branch a task and parent-work reference.
3. Classify capacity at a high level: exploration, execution, verification, recovery, coordination, or rework.
4. Record only minimal aggregate usage and an outcome summary.
5. Review one expensive successful run and one expensive unsuccessful run for attribution gaps.
6. Use the findings to improve task decomposition, reuse, stopping rules, and verification reserves—not to create a simplistic productivity leaderboard.

The goal is not to justify every token. It is to make resource decisions legible enough that a team can spend more when the user value warrants it, spend less when work is duplicate, and preserve the verification that makes an agent's outcome trustworthy.

The Virillio Code editorial perspective

The general lesson is that a budget controls what an agent may spend, while attribution explains what it spent and what task value or uncertainty reduction resulted. The Agent Cost Attribution Record is a proposed educational framework, not a description of a shipped or planned Virillio Code capability.

Sources and further reading

• Usage — OpenAI Agents SDK — The SDK documents aggregate and per-request usage counts and task-oriented usage data that can support detailed cost calculation and tracing.
• Gen AI semantic conventions — OpenTelemetry — The registry records GenAI usage and tool observability attributes, directs moved definitions to the dedicated GenAI conventions repository, and warns that content-bearing attributes may contain sensitive information.
• CostBench: Evaluating Multi-Turn Cost-Optimal Planning and Adaptation in Dynamic Environments for LLM Tool-Use Agents — A clearly labeled preprint studies cost-aware tool planning in a travel-planning environment with different tool costs and dynamic blocking events such as failures and cost changes.
• How we built our multi-agent research system — Anthropic — An engineering account describes multi-agent capacity costs and the need to reserve added effort for work whose value warrants it.

Editorial disclosure

This article was substantially researched, drafted, and revised with AI through the Virillio Code editorial workflow. Virillio Code publishes the final text under its editorial byline, and the supporting primary sources are linked above.