Agent Cost in Production

1type Budget = { maxIterations: number; maxCostUnits: number; tenantId: string }
2 
3async function runAgent(input: Input, budget: Budget): Promise<Result> {
4  let iterations = 0
5  let spent = 0
6  const trail: Step[] = []
7 
8  while (true) {
9    // Checked before every step, not once at request entry.
10    if (iterations >= budget.maxIterations || spent >= budget.maxCostUnits) {
11      emit('agent.budget_exhausted', { tenantId: budget.tenantId, iterations, spent, reason: iterations >= budget.maxIterations ? 'iterations' : 'cost' })
12      return degrade(trail)          // partial result, escalation, or handoff — never a silent truncation
13    }
14    if (!await tenantBudget.reserve(budget.tenantId, ESTIMATED_STEP_COST)) {
15      emit('agent.tenant_budget_exhausted', { tenantId: budget.tenantId })
16      return degrade(trail)
17    }
18 
19    const step = await model.step(trail, input)
20    spent += step.costUnits             // measured from the response, not estimated
21    iterations += 1
22    trail.push(step)
23 
24    if (step.kind === 'answer') return complete(trail)
25    if (!toolGuard.allows(step.tool, budget.tenantId)) {   // same guard as the kill switch
26      trail.push(unavailable(step.tool))
27      continue
28    }
29    const result = await callTool(step.tool, step.args)
30    spent += result.costUnits           // tool cost counts against the same budget
31    trail.push(result)
32  }
33}

Three things carry the lesson. The check is inside the loop, so an in-flight request is bounded. Tool cost counts against the same budget as model cost, because a loop over an expensive internal API is the same problem. And exhaustion calls degrade, which returns a marked partial result or a handoff — a truncated answer presented as a complete one is a quality regression that will be blamed on the model (In-the-Loop vs On-the-Loop and Escalation).