429 Too Many Requests (OpenRouter)
OpenRouter 429s come from three different places: free-model daily caps, your account's request-rate allowance (tied to credit balance), or the upstream provider itself throttling — and the error looks the same for all three. Fixes differ per cause; the structural fix for agents is an endpoint without per-minute economics.
Agents are disproportionately hit because retries-on-429 create feedback loops: each rejected request triggers a retry burst that keeps you over the limit. Cap retry concurrency to 1 with exponential backoff as the immediate mitigation.
Standard Compute serves agent workloads without per-minute rate economics: sustained load is paced smoothly rather than rejected with 429s, and smart routing spreads steps across the model pool — one flat monthly price, same OpenAI-compatible integration. If you like OpenRouter's model pinning, keep it for experiments; move the daily agent loop to the endpoint that doesn't meter it.
Request-rate allowance and credit balance are related but separate: you can have credits left and still exceed requests-per-minute, especially with an agent making parallel calls. The other common case is the upstream provider throttling — that 429 isn't OpenRouter's at all.
Only when the same model is served by multiple providers and the alternative has headroom. Single-provider models and account-level allowances aren't helped by fallbacks.