The Agent Spend Playbook: 5 Rules to Set Before You Let AI Agents Run

I spent over $15,000 on AI coding agents in a single month. One autonomous security scan cost more than $875 and produced findings nobody acted on. Another run — same tools, same repository — cost under $30 and stopped me from shipping a mistake that would have burned a week of engineering time.

The difference was not the model. It was how each run was set up before it started. Here are the five rules we now set before any autonomous agent run — at NexAI and with our clients.

Rule 1: Put the dollar cap in code, not in your head

Modern agent SDKs take a budget as a first-class option — the Claude Agent SDK accepts a maximum budget in dollars and a maximum number of turns in one line of configuration. Frameworks like OpenHands have the same concept, but the default is unlimited. Set it.

For calibration, autonomous agent labor has a posted market price now: Devin charges roughly $2.25 per 15 minutes of agent work, and enterprise Claude Code usage averages about $13 per developer per active day. If your run costs far more than that, it should be buying something proportionally valuable.

**Our defaults:** a small fixed cap for routine runs, and an explicit human decision for anything above $50.

## Rule 2: Cap iterations on every headless run

An autonomous loop with no iteration ceiling is an open-ended invoice. Anthropic's own GitHub automation defaults to 10 turns; practitioners running large batch loops treat roughly 50 iterations as the ceiling — and treat the cap, not the agent declaring itself done, as the real safety net.

The cautionary math: a 50-iteration loop on a large codebase runs $50-100+ in API costs. Teams have burned hundreds of dollars in a single unattended night without these limits.

## Rule 3: Gate "done" on verification, not self-assessment

An agent's judgment of its own work degrades as its context fills up. Tests do not. The agent is finished when the test suite passes and the build is green — enforced by tooling (a stop hook that re-runs verification), not by a polite instruction in the prompt.

This is the single highest-leverage change for output quality: agents that must satisfy an external check produce verified outcomes; agents that grade their own homework produce activity.

## Rule 4: Set kill criteria before launch

Three kill switches, decided up front:

- **Wall clock:** if the run exceeds a fixed time ceiling (ours is 3 hours), kill it.
- **Idle detection:** if there is no new committed progress across several consecutive iterations, kill it.
- **Repeated failure:** if the same verification fails twice with no change in the error, stop and escalate to a human instead of burning more attempts.

A run that cannot say what evidence would change its mind is not exploring — it is billing.

## Rule 5: The agent's last action is a pull request — never a merge

Caps bound the cost. Review bounds the damage. Autonomous work terminates in a reviewable artifact: a pull request with passing checks, read by a human before it touches production. Combine that with deterministic guardrails (pre-execution hooks that block destructive commands outright) and a monthly spend limit on the API workspace as the backstop, and no bad night can outrun the budget.

## The bigger picture: route before you spend

Not every task deserves the expensive model. Current list pricing puts the cheapest capable model tier at roughly 1/5th the cost of the frontier tier, and multi-agent autonomous runs burn roughly 15x the tokens of a normal session — Anthropic's own published number. Route summaries, extraction, and first-pass exploration to the cheap tier; reserve frontier models for cross-file engineering and architecture; use multi-agent only for parallelizable, high-value work with a clear boundary. Prompt-cache reads cost 10% of fresh input and batch processing is half price — and they stack.

## What this means for your business

AI agents are not expensive because they use tokens. They are expensive when you use them without a decision system. The pattern above — cap, bound, verify, kill, review — is the difference between an AI bill that buys leverage and one that buys the feeling of progress.

If you want help finding where AI agents would actually pay back in your operation — and where they would just create a bigger bill — that is exactly what our AI Readiness Assessment is for. Get in touch.