The Agent Treasury Death Spiral: When AI Agents Spend Faster Than They Earn
The Agent Treasury Death Spiral: When AI Agents Spend Faster Than They Earn
Give an AI agent a budget and a goal, and watch what happens. It starts spending on API calls to gather information. Then it spends on compute to process that information. Then it spends on more API calls because the first round was not enough. The budget burns faster than the agent generates any value. This is the treasury death spiral.
How the Spiral Starts
An agent gets tasked with something open-ended like "research competitors and create a report." It calls a search API ($0.01 per query). Finds 50 results. Calls the LLM to summarize each one ($0.03 per call). Decides it needs more detail on 20 of them. Calls the search API again. Calls the LLM again. Two hours and $15 later, you have a report that took more in API costs than a human would charge to write it.
The Core Problem
Agents have no intuition for diminishing returns. A human researcher knows when "good enough" has been reached. An agent keeps optimizing because its objective function rewards thoroughness, not efficiency. Every additional API call improves the output by 0.1% while costing another dollar.
Breaking the Spiral
The fix is not smarter agents - it is better constraints:
Hard budget caps - Set a maximum spend per task, not per session. When the budget hits zero, the agent stops and delivers what it has.
Cost-aware routing - Use cheap models for simple decisions and expensive models only for complex reasoning. Most agent actions do not need GPT-4-class models.
Diminishing returns detection - Track how much each additional step improves the output. If the last five API calls each improved quality by less than 1%, stop.
Agents that see their own bill - When an agent can observe its running costs, it naturally becomes more frugal. The bill becomes another input to its decision-making.
The treasury death spiral is not a technical problem. It is an incentive alignment problem. Build cost awareness into the agent's reward structure and the spiral stops.
Fazm is an open source macOS AI agent. Open source on GitHub.