On October 23, 2026, AWS introduced Infinity, an early research runtime from Hydro, its collaboration with UC Berkeley, under Apache 2.0. It saves the conversation at a tool call, suspends the agent and resumes it when a result arrives. The “75,000 agents in 8 GB” figure estimates memory from a simulated-model benchmark on AMD EPYC. It does not demonstrate that many language models executing simultaneously on a Raspberry Pi.

Waiting is a separate resource problem

An AI task can be short as an instruction and lengthy in execution. Consider an assistant waiting for a build to finish before summarizing its results. There is no reason to request another model response continuously during the build. The useful question is how to retain enough state to continue correctly when the event arrives.

Our editorial interpretation is that such architecture should be evaluated across its complete lifecycle. Avoiding idle execution does not eliminate the cost of active generation, storage or event delivery. A small inactive-conversation footprint does not determine how many simultaneous requests the real model or its provider can handle.

A reliable-resumption example

For a pilot, we would use one background test followed by a report. The agent dispatches the test, the environment restarts, and the result arrives afterward. Success would mean that the correct conversation receives that result and the report does not present an old execution as a new one. This is a proposed experiment that SIExpo has not performed.

The next cases are more revealing than the happy path: a duplicate event, a late result after cancellation, and results for two different tasks. For an operation with an external effect, the system should distinguish resuming reasoning from repeating the action itself. Otherwise, inexpensive waiting could end in a duplicated operation.

We would record the task identifier, saved-state version and time at which a result was accepted. Separately, we would check that a cancelled user intention remains cancelled after a restart. A conversation may be technically recoverable while no longer carrying permission to continue its original operation.

Those records should let an operator explain why the agent resumed and which pending operation it believed it was completing. Without that connection, a queue can deliver messages successfully while the overall workflow still reaches the wrong conclusion.

Reading the scaling claim

A useful comparison would measure memory while waiting, load when many events arrive, and end-to-end task completion time. These quantities answer different questions. If all events arrive together, the bottleneck may move to the model, tool or store, regardless of how inexpensive the preceding wait was.

Infinity offers a useful starting point for teams handling long-running workflows and interruptions. Its early research status calls for an owned assessment of recovery and permissions. The substantive development is the execution model based on saved state; the memory figure should remain attached to the conditions that produced it.

*Cover: conceptual AI illustration, not a documentary photograph.*

*Editorial verification: 2026-10-25T07:09:22.750908+00:00.*