Rate‑Limiting AI Agent Egress to Already‑Allowed APIs

Описание и применение

Проблема: список разрешенных адресов не ограничивает частоту запросов AI‑агентов, что приводит к перегрузке API. Решение: применить локальный лимит запросов на маршруте агент‑шлюза, ограничивающий трафик до 5 запросов в секунду после 20‑запросного всплеска.

Задача

An egress allowlist does not control how often an AI agent can hit an allowed API, leading to excessive traffic that can overwhelm the target.

Результат

The guarded route limited traffic to the configured burst and steady‑state rate, while the unguarded route allowed all traffic, demonstrating that a route‑level rate limit effectively protects an allowed API from agent overuse.

Рабочий процесс

• 1. Configure agentgateway with a localRateLimit policy (maxTokens:20, tokensPerFill:5, fillInterval:1s, type:requests).
• 2. Deploy two gateway instances: one guarded, one unguarded.
• 3. Send 60 requests via curl to each route.
• 4. Verify guarded route returns 429 for excess requests and upstream receives only 20 requests.
• 5. Verify unguarded route forwards all 60 requests to upstream.

Доказательства

An egress allowlist decides whether an agent can reach a destination. It has no opinion about how often.

Источник

The guarded route: localRateLimit { maxTokens: 20, tokensPerFill: 5, fillInterval: 1s, type: requests } Same 40 requests, same speed: 20 x 200, 20 x 429

Источник

scripts/verify.sh asserts the claim directly rather than eyeballing the demo output: it fires 60 requests through each route and checks that the unguarded upstream received all 60, the guarded upstream received fewer, and only the guarded route ever returned a 429.

Источник