Why your AI agent keeps getting rate limited (and how to stop it)
AI agent keeps getting rate limited
Agents hit rate limits far more than chat apps because they make requests in the background (heartbeats, retries, parallel tool calls and sub-agents) that quietly burn your per-minute provider budget. Cut that background load and add backoff, or move the agent to a provider with more headroom per key: Standard Compute allows 300 to 2,000 requests per minute per API key depending on the plan ($19 Starter to $249 Pro), and retries a provider's 429 up to twice, letting it fall back to another provider for the same model, before your agent sees an error.
What causes it
- Heartbeat / polling loops making model calls on a timer, even when idle.
- Retry-on-failure logic that multiplies calls during a provider hiccup.
- Parallel tool calls and sub-agents fanning out many requests per step.
- Re-sending large contexts every step, blowing tokens-per-minute limits.
How to fix it
- Lower heartbeat frequency and cap concurrency so background load stays under the limit.
- Add exponential backoff so retries don’t become a storm.
- Configure fallback providers so a throttled one doesn’t stall the agent.
- Compact context between steps instead of resending the whole history.
Running an agent? This is the #1 reason OpenClaw and Hermes agents stall. The durable fix is a per-key ceiling well above what the agent sends, plus retries that happen before the agent sees the error.
Rate limits by provider
LLM API rate limits compared →
OpenAI, Anthropic and Gemini tiers side by side, and how each one handles a 429.
Keep working on a fixed monthly plan
Standard Compute is built for always-on agents: each API key gets 300 (Starter), 600 (Economy), 1,000 (Standard) or 2,000 (Pro) requests per minute, provider 429s are retried and can fall back to another provider for the same model, and the price is one flat monthly plan with no 5-hour or weekly windows.
Plans from $19 a month with $19 of compute. Eligible new accounts can try $0.25 of compute first, no card. Trial terms
FAQ
Why do agents get rate limited more than normal apps?
Because they generate requests on their own (heartbeats, retries, and tool calls), independent of user activity. That background traffic competes for the same per-minute budget and trips limits you’d never hit by hand.
Is upgrading my provider tier enough?
It buys headroom, but a busy agent grows into the new ceiling. Cap concurrency in the agent and pick a provider whose per-key ceiling sits well above what the agent sends. On Standard Compute that is 300 requests a minute on Starter up to 2,000 on Pro, and 2,500 to 16,000 per 10 minutes.
What does Standard Compute return when a key hits its request ceiling?
HTTP 429 with the code rate_limit_exceeded, a Retry-After header (60 or 600 seconds, depending on which window was hit) and a message saying this is the plan's request-rate ceiling, not the compute budget. Running out of the monthly budget is a different response: HTTP 402 with the code budget_exhausted.