Set quotas per agent grant

How to do it

Agents call far faster than people and retry in loops. Without a limit per grant, one runaway agent uses up the quota that the user's other agents and the user's own browser need. Keep three counters, for the grant, the user and the verified operator from Set a policy for each agent operator. Refuse when any of them is over. Let the user set each agent's share in the account page, with a default such as half the user's quota. Answer an agent over its limit with 429, Retry-After and the headers in Tell clients their rate limit. Tenant-wide fairness is in Give each customer a fair share.

Examples

JavaScript example

export async function withinQuota(redis, { userId, agentId, operator }, limits) {
  const minute = Math.floor(Date.now() / 60_000);
  const keys = [`q:grant:${userId}:${agentId}:${minute}`, `q:user:${userId}:${minute}`, `q:op:${operator}:${minute}`];
  const counts = await Promise.all(keys.map(async key => { const n = await redis.incr(key); await redis.expire(key, 120); return n; }));
  return counts[0] <= limits.grant && counts[1] <= limits.user && counts[2] <= limits.operator;
}

Python example

def within_quota(redis, user_id: str, agent_id: str, operator: str, limits: dict) -> bool:
    minute = int(time.time() // 60)
    keys = [f"q:grant:{user_id}:{agent_id}:{minute}", f"q:user:{user_id}:{minute}", f"q:op:{operator}:{minute}"]
    pipe = redis.pipeline()
    for key in keys:
        pipe.incr(key)
        pipe.expire(key, 120)
    grant, _, user, _, op, _ = pipe.execute()
    return grant <= limits["grant"] and user <= limits["user"] and op <= limits["operator"]

How to check

A test gives one user two agent grants, each with half the quota. The first agent sends calls until it gets 429. The second agent's next call and the user's own next call both succeed.

Sources

• OWASP Top 10 for LLM Applications, LLM10 Unbounded Consumption
• MCP specification on tools (security considerations)

Related rules

• AGT-3 Set a policy for each agent operator in Agents as users
• API-36 Tell clients their rate limit in API design
• SCL-31 Give each customer a fair share in Scaling