Your AI Assistant Just Bought a $30,000 Cloud Subscription
A postmortem of the $30K Claude bill incident.
The Story
In May 2026, a story made the rounds: "AWS user gets $30K Claude bill after cost alert misses it." Two weeks later, another company reported a $38,000 AWS Bedrock bill caused by a prompt caching miss.
A single prompt cache miss. $38,000. Not a billion-dollar enterprise. A regular business running AI agents.
How Runaway Costs Actually Happen
When you tell an AI agent to "research competitors and draft a report," here's the execution graph:
1. Search API -> $0.03
2. Web scrape -> $0.01
3. GPT-4 summary -> $0.35
4. Agent decides: "not polished enough"
5. GPT-4 premium -> $2.50
6. Image gen API -> $1.00
7. Regenerate x 3 -> $7.50
8. Total -> $13.39 for one report
An agent doesn't know the difference between a $0.01 action and a $10 action.
The Architecture of Prevention
A library can be monkey-patched. A proxy is a network boundary agents must cross.
agent-gov is a FastAPI reverse proxy:
# Your config changes from:
openai.base_url = "https://api.openai.com/v1"
# To:
openai.base_url = "http://localhost:8080/v1"
The proxy runs a 4-stage decision tree:
@app.post("/proxy/call")
async def proxy_tool_call(call: ToolCall):
key_hash = db.hash_key(call.agent_key)
agent = await db.get_agent(key_hash)
# Stage 1: Auth - known agent?
if agent is None:
raise HTTPException(status_code=401)
# Stage 2: Paused?
if agent["paused"]:
raise HTTPException(status_code=429)
# Stage 3: Lazy budget reset
agent = await db.check_and_reset_budget(agent)
# Stage 4: Real cost lookup (not agent's estimate)
registered_tool = await db.get_tool(call.tool_name)
actual_cost = registered_tool["cost_per_call"] if registered_tool else call.estimated_cost
# Stage 5: Budget check
if agent["spent_today"] + actual_cost > agent["daily_budget"]:
await db.pause_agent(key_hash)
raise HTTPException(status_code=429, detail="Budget exceeded")
# Approved
await db.update_agent_spend(key_hash, actual_cost)
await db.log_cost_event(key_hash, agent["name"], call.tool_name, actual_cost)
return {"status": "approved", "spent_today": updated["spent_today"]}
The Anti-Cheat: Tool Registry
If you trust the agent's estimated cost, an agent can claim GPT-4 costs $0.01 when it's $12.50.
agent-gov uses a tool registry:
registered_tool = await db.get_tool(call.tool_name)
actual_cost = registered_tool["cost_per_call"] if registered_tool else call.estimated_cost
# cost_source: "registry" or "client_estimate"
A test proves agents can't lie:
async def test_proxy_uses_registered_cost():
# Tool registered at Rs 500/call
# Agent with Rs 100 budget claims Rs 1
# Result: 429 - Blocked!
Why Proxy Wins Over Library
- Network boundary - agents must cross it
- Can't be bypassed by rogue import or version bump
- Language-agnostic - works with any framework
- Externally monitorable
Quick Start
pip install agent-gov-saas
agent-gov start
agent-gov config set budget 25.00 --agent my-bot
Auto-paused at $25. No $30K surprise.
Part 1 of "Taming Your AI" series. agent-gov is open-source, MIT-licensed.