How to Limit OpenAI API Spend Per Agent Run

[bmdhodl]

If you're running AI agents in production, you've probably wondered: how do I stop a single agent run from spending more than $X?

OpenAI's dashboard has organization-level spend limits, but those are blunt — they shut down everything, not just the one runaway agent. And they're reactive: by the time the limit kicks in, the money is gone.

Here's how to enforce a hard dollar cap per agent run using AgentGuard's BudgetGuard.

Quick setup (4 lines)

from agentguard import Tracer, BudgetGuard, patch_openai

tracer = Tracer(guards=[BudgetGuard(max_cost_usd=5.00)])
patch_openai(tracer)
# Every OpenAI call is now tracked. Agent dies at $5.

That's it. patch_openai wraps the OpenAI client so every chat.completions.create() call automatically:

1. Extracts token usage from the response
2. Estimates cost using published pricing (GPT-4o, GPT-4, GPT-3.5, etc.)
3. Checks against your budget
4. Raises BudgetExceeded if the limit is hit

Warning before the hard stop

You probably want a chance to wrap up gracefully before the kill:

guard = BudgetGuard(
    max_cost_usd=5.00,
    warn_at_pct=0.8,  # Warning at 80% ($4.00)
)

Checking spend mid-run

guard = BudgetGuard(max_cost_usd=10.00)

# After some calls...
print(f"Spent: ${guard.state.cost_used:.4f}")
print(f"Calls: {guard.state.call_count}")
print(f"Remaining: ${10.00 - guard.state.cost_used:.4f}")

Also works with Anthropic

from agentguard import patch_anthropic

patch_anthropic(tracer)
# Claude calls are now tracked too. Same BudgetGuard, same limits.

Install

pip install agentguard47

Zero dependencies, MIT licensed, Python 3.9+.

GitHub: https://github.com/bmdhodl/agent47

---

What's your current approach to capping per-agent spend? Have you hit any edge cases with token estimation?