Measure cost per verified outcome, not cost per response

[
  {
    "field": "task_budget",
    "purpose": "Maximum total spend, elapsed time, model calls, and tool calls for one task.",
    "evidence": "Enforced hard limit and terminal reason when exceeded."
  },
  {
    "field": "step_budget",
    "purpose": "Per-step limits that prevent one operation from consuming the entire task envelope.",
    "evidence": "Step counters, timeout, payload, and cost limit tests."
  },
  {
    "field": "delegation_budget",
    "purpose": "Maximum depth, fan-out, peer count, and delegated authority.",
    "evidence": "Non-amplifying delegation checks and blocked fan-out cases."
  },
  {
    "field": "communication_budget",
    "purpose": "Bounds messages, tokens, rounds, shared-state writes, and coordination overhead.",
    "evidence": "Communication totals and termination on non-progressing exchanges."
  },
  {
    "field": "retrieval_budget",
    "purpose": "Bounds sources, bytes, queries, and freshness checks.",
    "evidence": "Coverage and quality remain acceptable within the limit."
  },
  {
    "field": "tool_budget",
    "purpose": "Bounds calls by tool, effect class, tenant, value, and environment.",
    "evidence": "Denied over-limit action and reconciled unknown result."
  },
  {
    "field": "retry_budget",
    "purpose": "Prevents unbounded retries and retry storms across nested agents.",
    "evidence": "Reason-aware retry count, backoff, idempotency, and stop case."
  },
  {
    "field": "human_review_budget",
    "purpose": "Makes reviewer capacity, queue delay, and escalation cost visible.",
    "evidence": "Assignment delay, review time, expiry, and overload behavior."
  },
  {
    "field": "context_budget",
    "purpose": "Limits prompt, memory, retrieval, and artifact context to reduce cost and data exposure.",
    "evidence": "Context selection record and quality regression test."
  },
  {
    "field": "provider_budget",
    "purpose": "Limits concentration and spend by model, tool, or infrastructure provider.",
    "evidence": "Provider allocation, fallback behavior, and concentration threshold."
  },
  {
    "field": "outcome_cost",
    "purpose": "Measures cost per verified outcome rather than cost per model response.",
    "evidence": "Task, effect, human review, recovery, and failed-attempt cost included."
  },
  {
    "field": "change_trigger",
    "purpose": "Reopens evaluation when price, model, tool, topology, or workload changes the budget envelope.",
    "evidence": "Versioned baseline, threshold, owner, and review date."
  }
]

Define the verified outcome

Name what counts as success: reviewed recommendation, accepted artifact, completed durable action, reconciled transaction, or another outcome. A generated answer is not the same as a verified effect or an accepted business decision.

Associate cost with the outcome and with failed, escalated, reversed, or partially completed attempts. Otherwise the unit economics will hide operational waste.

Create nested budget envelopes

Use a total task budget and smaller step, retrieval, tool, retry, delegation, communication, and human-review envelopes. The sum of lower-level allocations should not silently exceed the task's approved maximum.

Expose remaining time, cost, calls, and consequence capacity to deterministic orchestration. A model may propose how to use the remaining budget, but hard enforcement belongs outside the model.

Bound delegation and coordination

Limit depth, fan-out, peer count, rounds, messages, shared-state writes, and delegated authority. Record why a new agent is added and whether it contributes specialization, independent challenge, parallelism, or a necessary ownership boundary.

Compare total outcome quality, cost, latency, failure rate, and reviewer burden with a strong single-agent or deterministic baseline. Research on multi-agent systems shows that additional agents can introduce coordination failures and may underperform the strongest member. [S1] [S2]

Control retries and unknown completion

Retries need a reason, attempt limit, backoff, idempotency key, and status reconciliation. Tool timeout can mean rejected, failed, completed, partially completed, or unknown. Blind retry can duplicate an external effect and multiply cost.

Track nested retry amplification when one parent agent retries a workflow whose child agents also retry independently.

Include data, tool, and provider costs

Measure retrieval queries and bytes, embedding and reranking, vector or database operations, tool/API charges, network egress, storage, event processing, and provider minimums. Track provider concentration and fallback cost.

Price or quota changes, model migrations, new tools, and larger context windows are material changes that can reopen the release and commercial decision.

Make human review visible

Include assignment delay, review minutes, specialist involvement, rework, disagreement, override, escalation, and exception handling. A low model cost can be uneconomic when every case requires lengthy expert review.

Set queue and expiry limits so a human checkpoint does not become an invisible unbounded backlog.

Measure tail behavior and capacity

Report percentiles and worst-case bounded behavior for latency, calls, cost, messages, retries, and reviewer time. Average cost can hide rare loops that dominate spend or create service instability.

Run load and scarcity tests. Observe whether the system degrades predictably, sheds work, escalates, or continues consuming resources without progress.

Use budgets as safety controls

Budget exhaustion should produce a safe terminal, partial-result, or escalation state with evidence and no unauthorized side effect. Do not let a truncated context or forced stop silently present an incomplete result as complete.

For high-consequence tools, combine resource budgets with amount, tenant, environment, and effect-class limits.

Review economics after release

Compare forecast and actual cost per verified outcome, outcome acceptance, error correction, reviewer burden, incident cost, and value created. Reopen the architecture when coordination overhead or recovery cost erases the measured benefit.

This article does not provide a universal price, ROI claim, or expected-revenue model. Economics must be measured in the actual workflow.