Soft Alerts Don’t Stop Agent Spend. Hard Budgets Do.

Soft alerts are comforting.
They are also how you wake up to a surprise OpenAI bill from a retry loop at 3am.

If you run multi-step agents (tools + LLM calls), you need two systems:

1. Observability — what happened (traces, tokens, errors)
2. Enforcement — what is allowed to happen next (budget, tools, memory scope)

Most stacks are strong at (1) and weak at (2).

---

Soft alerts vs hard budgets

Approach	What it does	Failure mode
Alert at 80% of monthly spend	Email / Slack	Agent keeps running
Dashboard “cost this week”	Human looks later	Overnight burn
Hard budget on the run	Reject next spend when ceiling hits	Run stops; you debug

Hard budgets only work if LLM/tool cost is recorded on the hot path (or immediately after each call). A run with only “start/finish” spans cannot enforce real dollars.

---

Minimal mental model

Agent run starts with budget_usd = 0.50
→ tool call
→ LLM call (+ tokens → $)
→ if spent >= budget → fail closed (raise / stop)
→ else continue

Enforcement should not be “we'll audit later.”
If the gate is only after the wire transfer (or the refund email), it's already too late.

---

Python sketch (pattern)

You can implement this in app code. The important parts:

• Attach a budget to the run (not only org monthly total)
• Record every LLM call’s estimated cost
• On exceed: raise (or return a controlled error) — don’t log-and-continue

python
class BudgetExceeded(Exception):
    pass

class RunBudget:
    def __init__(self, limit_usd: float):
        self.limit = limit_usd
        self.spent = 0.0

    def charge(self, usd: float) -> None:
        if self.spent + usd > self.limit:
            raise BudgetExceeded(
                f"spent={self.spent:.4f} + {usd:.4f} > limit={self.limit}"
            )
        self.spent += usd

Wire that into your OpenAI wrapper after each completion (use usage tokens × your price table).