AI agents spend most of their time waiting: on a build that runs for twenty minutes, on the next pull request, or on a human approval that arrives the next morning. AWS and UC Berkeley have now open-sourced Infinity, a runtime built around that reality — instead of keeping agents running while they wait, it suspends them during tool calls and resumes them when results arrive.
The approach, similar to stackless coroutines in programming, delivers striking efficiency. Infinity supports 75,000 concurrent agents on 8 GB of memory — demonstrated on a Raspberry Pi 5 — with roughly 100 KB per idle agent. Because suspended agents consume no compute, the runtime also runs on serverless platforms such as AWS Lambda with SQS and Aurora DSQL.
Infinity handles the awkward parts of long-running agents too. It uses synthetic tool calls to deliver asynchronous events within model API constraints, processes events in child threads by default to avoid bloating the parent agent’s context window, and enables agent-to-agent communication through built-in spawn_thread and report_to_parent tools.
For teams building agent systems with tools like Claude Code, Codex or Cursor, the runtime matters because the bottleneck in production agents is rarely model intelligence — it is cost and concurrency at scale. Long-running, event-driven agents that once needed a dedicated process each can now idle almost for free.
By treating tool calls as suspension points, Infinity reframes agent infrastructure around waiting rather than thinking — and that may be the change that makes fleets of thousands of agents affordable.
Source: AWS News Feed
