AWS open-sources Infinity, a runtime for massively concurrent, long-waiting agents

AWS's open-source blog introduced Infinity, a runtime aimed at a problem most agent frameworks ignore: agents mostly wait. A coding agent might kick off a twenty-minute build, triage a pull request, or wait overnight for a human approval. Keeping a full process and context resident during those waits wastes compute, and that waste grows linearly with the number of agents.

Infinity's design does not keep each agent loop resident. Agents can suspend while blocked on tools or approvals and resume when an event arrives. That lets operators run very large numbers of concurrent, long-lived agents without paying for idle memory and CPU. The model resembles durable workflow engines applied to LLM agent loops: state is checkpointed, and execution wakes on external signals.

Infinity sits alongside a busy week of agent infrastructure. AWS also updated AgentCore Gateway with fallback routing, Google introduced Agent Compute Engine for isolated concurrent sessions, and NVIDIA promoted OpenShell for agent fleets. Each vendor is betting that the bottleneck has moved from model quality to the plumbing that runs thousands of agents reliably and cheaply. Open-sourcing Infinity also positions AWS for multi-cloud adoption rather than lock-in to AgentCore alone.

The caveats are practical. Suspend/resume runtimes only pay off if agents' state is serializable and tools are idempotent. Developers will want to see how Infinity handles in-flight tool calls and context restoration, and whether snapshot restores lose context — a concern already raised about Google's microVM approach. Benchmarks and integrations with popular agent SDKs will decide uptake.