How to Limit OpenAI API Spend Per Agent Run

[bmdhodl]

How to Limit OpenAI API Spend Per Agent Run

Category: Show and tell
Labels: openai, budget, cost-control

---

If you're running autonomous agents with OpenAI, you've probably wondered: how do I set a hard dollar limit per agent run?

OpenAI has account-level spend limits, but those are a kill switch for your entire organization. They don't help when one agent run out of 50 goes haywire.

Here's how to set a per-run budget that kills the agent the moment it exceeds your limit.

Setup

pip install agentguard47

from agentguard import Tracer, BudgetGuard, BudgetExceeded, patch_openai

tracer = Tracer(
    service="my-agent",
    guards=[BudgetGuard(max_cost_usd=2.00, warn_at_pct=0.8)],
)
patch_openai(tracer)

# Now use OpenAI normally:
import openai
client = openai.OpenAI()

try:
    for i in range(100):
        response = client.chat.completions.create(
            model="gpt-4o",
            messages=[{"role": "user", "content": f"Step {i}: analyze data"}],
        )
except BudgetExceeded as e:
    print(f"Agent stopped: {e}")
    # Agent killed at $2.00 — no more API calls

patch_openai intercepts every ChatCompletion response, extracts token counts, estimates cost using published pricing, and feeds it into BudgetGuard. When the cumulative cost exceeds your limit, BudgetExceeded is raised.

Warning before the hard stop

The warn_at_pct=0.8 parameter fires a callback at 80% of the budget. You can use this for graceful shutdown:

guard = BudgetGuard(
    max_cost_usd=5.00,
    warn_at_pct=0.8,
    on_warning=lambda msg: print(f"WARNING: {msg}"),
)

Zero dependencies

AgentGuard is pure Python stdlib. No extra packages to audit.

pip install agentguard47

Repo: github.com/bmdhodl/agent47

---

What's your current approach to limiting per-agent spend? Curious if anyone has built custom solutions.