On October 23, 2026, AWS introduced Infinity, an agent runtime built with UC Berkeley, in its open source blog: instead of treating an agent as a resident process that always occupies a machine, it treats a tool call as a suspension point. When an agent makes a tool call, it suspends and saves its conversation history, then resumes once the result arrives, an approach similar to stackless coroutines in programming. The change targets the most expensive side of agents: much of their time is spent waiting on tools, approvals, or external events, not reasoning.
What suspension actually saves
In the conventional approach, an agent waiting for a result still holds a process and memory, so a crowd of idle tasks fills the machine first. Infinity packs away the waiting state, keeping only a lightweight record, and reconnects the conversation history and pending work on resume. For developers, the calling style barely changes; what changes is how the runtime treats waiting: waiting no longer means occupying resources.
The official number is striking: on a Raspberry Pi 5 with 8GB of memory, Infinity can carry 75,000 concurrent agents, with each idle agent taking about 100KB. Two bills must still be kept apart: that figure describes how many agents can be suspended at once, not how fast they all run, and the cost and latency of model calls remain.
It also runs serverless
Infinity can run in a local process or on a serverless platform such as Lambda, together with SQS and Aurora DSQL: while an agent waits, it needs no compute, and it wakes when the result arrives. That suits long-running, event-driven work, such as waiting for a build to finish, a human approval, or a reply from an external system, without renting machines long-term just for waiting.
Multi-agent collaboration is built into the runtime as well. Infinity provides built-in tools such as spawn_thread and report_to_parent, letting a parent agent spawn child threads and receive their results back; events are handled in child threads by default, so intermediate results are not all stuffed into the parent agent's context. When a model API does not itself support asynchronous events, Infinity delivers them with synthetic tool calls.
Who benefits first, and what to watch
The first beneficiaries are teams with many agents and long waits: customer service, monitoring, batch processing, and long follow-up tasks no longer need a whole fleet of resident capacity reserved for peak concurrency. For individual developers, the Raspberry Pi-scale demo also lowers the cost of trying it out.
Three cautions remain: suspension saves waiting-time compute, not model call fees; with more asynchronous flows, a stuck child thread or a late event makes debugging feel more like tracing a distributed task; and an agent's judgment quality, tool permissions, and safety boundaries still depend on the design and review above the runtime. It is more accurate to see Infinity as infrastructure that makes agents affordable to wait with, rather than as agents becoming smarter.
