Watch Token Costs

Agent loops are invisible money fires. A task that looks free can cost $5. A task that looks cheap can cost $50. Cost awareness is part of the operator's job.

Frontier models charge per token. Agents send context back and forth repeatedly, often with large code files attached. Without a budget, a single debugging session can exceed a restaurant dinner in API spend.

1. Set a daily or weekly budget. Use provider dashboards or tools like Langfuse to track spend.
2. Use cheaper models first. Start with a fast, cheap model. Escalate to an expensive one only when needed.
3. Limit context. Don't paste entire codebases into every prompt. Send only relevant files.
4. Cap retries. A stuck agent can loop forever. Set max iteration counts.
5. Review weekly. Look for spikes and patterns. One expensive workflow is worth optimizing.

• You don't know what you spent this week.
• Every task uses the most expensive model by default.
• The agent re-reads the same large files in every turn.
• You are surprised by monthly bills.