Meta built an internal leaderboard called "Claudeonomics." It tracked AI token consumption across 85,000 employees. Gamified tiers from bronze to emerald. Titles like "Token Legend" and "Session Immortal." A competitive race to use the most AI.
In 30 days, they burned 60 trillion tokens.
Then they shut it down.
What happened
The Claudeonomics dashboard was a voluntary internal tool on Meta's intranet. It ranked the top 250 AI token consumers with gamified incentives. The idea was to encourage AI adoption across the company.
It worked too well.
Multiple sources confirmed that employees left AI agents running for hours executing busywork research tasks specifically to climb the leaderboard. They consumed tokens while producing nothing of value.
The top individual consumer averaged 281 billion tokens per day. For a month straight.
Why it matters
Token consumption is an input metric. Not an output metric. Measuring productivity by tokens consumed is like measuring engineering quality by lines of code written.
Meta learned this the expensive way. But the lesson applies to every team running AI agents in production.
Here is the pattern:
- Team deploys AI agents
- No budget limits set
- Agents run autonomously (or employees run them to look productive)
- Token costs compound without anyone watching
- Someone notices a $50,000 cloud bill
Meta can absorb the cost. Your team probably cannot.
The math at your scale
Let's scale it down. Say you have 5 agents running production tasks. Each processes 100 requests per day. Average cost per request: $0.10.
That is $50/day. $1,500/month. Manageable.
Now one agent hits a retry loop. It fires 10,000 requests in an afternoon. That is $1,000 in one burst. No warning. No cap. Just a bill.
Or an agent starts looping through a research task with no termination condition. It runs all weekend. Monday morning, you have a $3,000 bill and a 2MB log file of circular reasoning.
This is not hypothetical. This is the default behavior of every agent framework that ships without budget controls.
What Meta should have done
Three things:
1. Budget limits per agent, per session
Every agent needs a hard cap. Not a soft warning. A hard stop.
from agentguard47 import init, BudgetGuard
init(
guards=[BudgetGuard(max_cost=10.00)]
)
When the budget hits $10, the agent stops. No negotiation. No override. The guard is deterministic. The agent cannot convince it to keep going.
2. Loop detection
Agents loop. It is what they do when they get stuck. Without detection, a loop runs until something external kills it (usually the credit card limit).
from agentguard47 import init, LoopGuard
init(
guards=[LoopGuard(max_iterations=100)]
)
100 iterations and done. If the agent has not solved the problem in 100 tries, iteration 101 is not going to help.
3. Kill switches
Sometimes you need to stop everything. Right now. Not "after the current batch finishes." Now.
AgentGuard's timeout guard gives you that:
from agentguard47 import init, TimeoutGuard
init(
guards=[TimeoutGuard(max_seconds=300)]
)
Five minutes. Then it is over. Combine all three for defense in depth.
The real lesson
Meta's Claudeonomics experiment failed because they measured the wrong thing. But the deeper failure was structural: 85,000 people running AI agents with no runtime budget controls.
The gamification just made the problem visible faster.
Every team running AI agents without budget limits is running the same experiment. You just do not have a leaderboard showing you the results.
Set your limits before you need them. Not after.
AgentGuard is an open-source Python SDK for AI agent runtime safety. Budget limits, loop detection, and kill switches. Zero dependencies. Local-first.
Related: AI Agent Cost and Pricing in 2026
Top comments (0)