Cost Control¶

Finance asks how much each team spent on AI agents last month, by model and by tool. The answer must be defensible, reproducible, and produced without instrumenting every SDK separately.

Symptoms¶

Provider bills are aggregate; the team that incurred the cost is invisible.
Agents spike cost overnight and nobody knows why until the bill arrives.
Setting a budget on one team has no effect because the limit lives in spreadsheets.
A new agent ships with no cost expectation and runs unchecked.

What Joch does¶

Cost is a first-class observability primitive. The Model Router and Tool Gateway record costUsd per call. The Budget controller enforces caps before they are exceeded.

Walkthrough¶

1. Set budgets¶

apiVersion: joch.dev/v1alpha1
kind: Budget
metadata:
  name: support-platform-monthly
spec:
  scope: { teams: [support-platform] }
  limits: { monthlyUsd: 5000, dailyUsd: 250, perRunUsd: 5 }
  enforcement:
    onSoftLimit: alert
    onHardLimit: deny
    softLimitPct: 0.8

joch apply -f budgets/support-platform-monthly.yaml

2. View attribution¶

joch cost by-team --since 30d
joch cost by-agent --since 7d
joch top models --by cost --since 30d
joch top tools  --by cost --since 30d

Every line item joins through agent, team, environment, model, tool, and time-window.

3. Investigate spikes¶

joch executions ls --over-cost 4 --since 24h
joch describe execution exec-20260510-014
joch trace exec-20260510-014

The trace surfaces the model calls and tool calls that drove the cost, with token counts and per-call cost.

4. Respond with policy¶

apiVersion: joch.dev/v1alpha1
kind: Policy
metadata: { name: research-cost-cap }
spec:
  appliesTo:
    agents: { selector: { matchLabels: { team: research } } }
  rules:
    - when: { cost.runUsd: ">3" }
      action: { deny: true, reason: research-cap-exceeded }
    - when: { model.provider: openai, model.tier: premium }
      require: { approval: human }

5. Optimize routes¶

apiVersion: joch.dev/v1alpha1
kind: ModelRoute
metadata: { name: research-default }
spec:
  requirements: { toolCalling: true, contextWindowMin: 200000 }
  strategy:
    primary: anthropic:claude-haiku
    fallback:
      - openai:gpt-5-thinking
  constraints: { maxCostUsdPerExecution: 2 }

Switching the primary in one place reroutes every agent that uses the route. No agent code changes.

Resources involved¶

Outcome¶

Finance gets defensible cost reports. Engineering gets enforceable caps before bills arrive. Optimizing cost is a route or budget change, not a code change.