AWS describes it as an early research release.
Infinity treats an agent as a sequence of bounded execution slices. In each slice, it loads conversation history from a pluggable store, runs one model completion, persists the output, and dispatches at most one tool call. Rather than wait on that call, the runtime suspends the agent; a later tool result or external event resumes it from stored state.
AWS says this design can run on serverless infrastructure because no function needs to remain active while an agent waits. Its reported synthetic benchmark put idle-agent memory use at about 100 KB and fit roughly 75,000 concurrent agents into the 8 GB memory of a Raspberry Pi 5. The post also describes a Lambda-oriented mode using SQS FIFO, Aurora DSQL, and DynamoDB.
The source does not establish equivalent performance with production models, real tool traffic, or customer workloads. It also does not present Infinity as a generally available managed AWS service.