Manage the Money

Before diving in, complete Three-Tier Setup so your budget controls map to real model-tier assignments rather than a flat cost estimate for all agents.

Here is the before and after: You scale to five parallel Opus agents and your bill for that day is $180. It is not a bug. It is the math: 10 million tokens per run, Opus pricing, five agents. Most teams hit this within two weeks of scaling background agents. The fix is not to run fewer agents. It is to route tasks to the right model tier before you dispatch.

Model routing by stakes is the core pattern. Use the most capable model where it actually matters: complex reasoning, architectural decisions, first-pass implementation of ambiguous requirements. Use cheaper models where the task is well-defined. Sonnet for review and synthesis costs about 5x less than Opus. Haiku for formatting and label generation costs 20x less. A Sonnet PR review run costs roughly 8% of an Opus implementation run. Route correctly and the savings pay for themselves in the first week.

Per-run budgets are the circuit breaker. Every background agent run should have a maximum token cap. If the agent hits the cap, it surfaces what it has and stops. Without a cap, an unexpected loop consumes unbounded resources. The cap is not a quality ceiling. It is a hard stop.

Real-time spend monitoring means you see cost as it accumulates, not in next month's invoice. Treat cost the same way you treat test failures: a signal that fires while the work is running, not after.

Paste this into Claude:

I want to set up cost management for my background agents. Help me: (1) Look at the tasks I run most frequently and classify each as high-stakes (needs Opus-tier), medium-stakes (Sonnet-tier), or low-stakes (Haiku-tier). My tasks are: [list 4-6 tasks you run regularly, e.g. 'implement a new API endpoint', 'review a PR for logic errors', 'update doc comments', 'format code output', 'generate test cases']. (2) For each tier, write the model routing rule I should add to my supervisor or skill definition. (3) Set a per-run token budget for each tier and explain why. (4) Show me what a simple cost log entry should look like: model, tokens in, tokens out, estimated cost, task ID.

What good looks like:

• You classified your regular tasks into at least two model tiers (not everything goes to the most capable model)
• Each classification has a specific reason tied to task complexity, not just cost
• You defined per-run token budgets for at least two tiers
• You have a cost log format with model, tokens, estimated cost, and task ID
• You can estimate your daily cost at current usage and identify which task type drives most of it