AWS ships default rate limits for AgentCore agent traffic

AWS added per-user, per-target rate limits to its AgentCore gateway on August 6, per AWS's blog post. The caps cover MCP targets, inference targets and HTTP passthrough. The defaults are worth copying even for gateways that aren't running on Bedrock. That matters for anyone running multiple tools off one gateway: a single user tops out at 20 RPM inside a 100 RPM Basic tier.

Limits scope two ways. Bind them to identity, either JWT claims like sub, role and azp, or IAM identities directly. Or bind them to what's being called: targetName, toolName or qualifiedModelId. A single gateway can throttle one noisy tool without slowing the rest of the traffic running through it.

Token limits apply only to inference targets, tracked in tokens per minute; connection limits are tracked separately, in connections per second. Basic tops out at 80,000 TPM and 50 CPS, while Advanced trades a lower 40,000 TPM ceiling for 150 CPS. The individual-user scope gets 20,000 TPM. Requests, tokens and connections are tracked separately, so a client can burn through its token budget before hitting a request ceiling.

The Advanced tier's token budget is lower than Basic's, despite triple the request and connection headroom. Teams should size AgentCore limits by tokens, not by tier name. AWS didn't say why the tiers split that way. The mismatch is the detail worth copying into your own limits, even if you're nowhere near Bedrock.