The Hydro project's Rust runtime suspends agents during tool calls so roughly 75,000 idle agents can fit in 8 GB, enabling cheap serverless agent fleets.
AWS and UC Berkeley released Infinity on October 23, 2026, an Apache 2.0 Rust runtime that treats agent tool calls like async await points—suspend, persist, resume—so thousands of agents can idle without holding memory or CPU. For software teams drowning in polling loops, the pitch is economic and architectural at once.
The blocking agent problem
Most agent harnesses today run a synchronous loop: model, tool, wait, repeat. When a build takes twenty minutes or a human approval takes overnight, that process stays allocated. Infinity ends the slice after dispatching a tool, stores conversation state, and wakes the agent when a message arrives on its queue.
Shadaj Laddad's AWS Open Source Blog post walked through the lifecycle: load history, one completion, at most one tool call, persist, exit. Tool results enqueue the next slice.
Numbers that caught Hacker News attention
Benchmarks cited roughly 100 KB of resident memory per idle agent, enabling on the order of 75,000 concurrent agents in 8 GB—Raspberry Pi territory. The test ran 80,000 agents through millions of completions on a single AMD EPYC thread with in-memory stores.
Production deployments will differ, but the order of magnitude challenges teams paying for always-on Kubernetes pods per customer workflow.
Serverless and scale-to-zero
Because suspended agents are pure data, Lambda functions can exit while tools run elsewhere. SQS and Aurora DSQL appear in reference architectures as event sources. You pay for inference slices, not for waiting.
That aligns with how businesses actually use agents: long horizons, sparse events, unpredictable humans in the loop.
Synthetic tool calls and model limits
Infinity delivers asynchronous events through synthetic tool calls so existing chat APIs can stay unchanged. Agent-to-agent messaging uses built-ins like spawn_thread and report_to_parent, with child threads processing events to keep parent context windows lean.
Framework authors should study this pattern before inventing proprietary message buses for every provider.
Hydro project context
Infinity ships under the Hydro collaboration between AWS and Berkeley. It is explicitly an early research release seeking feedback, not a GA managed service with an SLA—yet.
Still, open source here is a credibility move: teams can prototype without committing to a proprietary orchestrator.
Who should experiment first
SRE teams running triage bots, developer platforms firing CI agents, and customer success tools waiting on ticket updates all match the workload profile. If your agent spends more than half its lifetime blocked on I/O, Infinity's model is worth a spike.
Integration considerations
You will need pluggable conversation stores, durable queues, and idempotent tool implementations. Race conditions do not disappear; they move to message ordering guarantees per thread_id.
Observability must trace slice boundaries, not just HTTP requests.
Comparison with proprietary clouds
Google's Agent Compute Engine emphasizes credential isolation; Infinity emphasizes concurrency economics. Real stacks will combine both: cheap idle time plus safe execution.
Rust and operator skill
Rust adoption in infrastructure tooling continues to climb. Teams without Rust expertise can still consume Infinity via hosted patterns once vendors wrap it, similar to how many teams run Envoy without writing C++.
Limitations acknowledged
Authors noted simulations with static models and in-memory stores; real tool latency distributions and provider rate limits need more study. Expect rough edges in October 2026.
Getting started
The GitHub repository hydro-project/infinity includes a minimal AgentSystemBuilder example launching a thread that asks for a haiku about Rust—developer marketing at its finest.
Takeaway
Infinity does for agents what async runtimes did for web servers: stop tying up scarce resources while waiting on the world. On the day OpenAI talked about offline training cages, AWS talked about efficient waiting rooms. Both are infrastructure for the agent decade.
Additional context
Teams reviewing this story on October 24, 2026 should archive primary sources—vendor blogs, exchange filings, and regulator PDFs—because fast-moving AI and crypto news gets rewritten quickly on social feeds. Decision-makers need a paper trail when they explain strategy shifts to boards or investors weeks later.
Analysts also recommend separating signal from engagement bait. Headlines optimized for fear or hype rarely include base rates, confidence intervals, or scenario planning. Stress-test both bull and bear cases before changing roadmaps.
For international readers, time zones matter: U.S. evening announcements often land as Asian morning trading sessions open, amplifying volatility. Coordinate communications if your company spans regions.
Finally, pair news consumption with internal drills. Tabletop exercises for agent security, treasury policy, or SEO measurement catch gaps that headlines merely describe.
Industry reactions
Teams reviewing this story on October 24, 2026 should archive primary sources—vendor blogs, exchange filings, and regulator PDFs—because fast-moving AI and crypto news gets rewritten quickly on social feeds. Decision-makers need a paper trail when they explain strategy shifts to boards or investors weeks later.
Analysts also recommend separating signal from engagement bait. Headlines optimized for fear or hype rarely include base rates, confidence intervals, or scenario planning. Stress-test both bull and bear cases before changing roadmaps.
For international readers, time zones matter: U.S. evening announcements often land as Asian morning trading sessions open, amplifying volatility. Coordinate communications if your company spans regions.
Finally, pair news consumption with internal drills. Tabletop exercises for agent security, treasury policy, or SEO measurement catch gaps that headlines merely describe.
Extended analysis
Teams reviewing this story on October 24, 2026 should archive primary sources—vendor blogs, exchange filings, and regulator PDFs—because fast-moving AI and crypto news gets rewritten quickly on social feeds. Decision-makers need a paper trail when they explain strategy shifts to boards or investors weeks later.
Analysts also recommend separating signal from engagement bait. Headlines optimized for fear or hype rarely include base rates, confidence intervals, or scenario planning. Stress-test both bull and bear cases before changing roadmaps.
For international readers, time zones matter: U.S. evening announcements often land as Asian morning trading sessions open, amplifying volatility. Coordinate communications if your company spans regions.
Finally, pair news consumption with internal drills. Tabletop exercises for agent security, treasury policy, or SEO measurement catch gaps that headlines merely describe.
Extended analysis
Teams reviewing this story on October 24, 2026 should archive primary sources—vendor blogs, exchange filings, and regulator PDFs—because fast-moving AI and crypto news gets rewritten quickly on social feeds. Decision-makers need a paper trail when they explain strategy shifts to boards or investors weeks later.
Analysts also recommend separating signal from engagement bait. Headlines optimized for fear or hype rarely include base rates, confidence intervals, or scenario planning. Stress-test both bull and bear cases before changing roadmaps.
For international readers, time zones matter: U.S. evening announcements often land as Asian morning trading sessions open, amplifying volatility. Coordinate communications if your company spans regions.
Finally, pair news consumption with internal drills. Tabletop exercises for agent security, treasury policy, or SEO measurement catch gaps that headlines merely describe.
Extended analysis
Teams reviewing this story on October 24, 2026 should archive primary sources—vendor blogs, exchange filings, and regulator PDFs—because fast-moving AI and crypto news gets rewritten quickly on social feeds. Decision-makers need a paper trail when they explain strategy shifts to boards or investors weeks later.
Analysts also recommend separating signal from engagement bait. Headlines optimized for fear or hype rarely include base rates, confidence intervals, or scenario planning. Stress-test both bull and bear cases before changing roadmaps.
For international readers, time zones matter: U.S. evening announcements often land as Asian morning trading sessions open, amplifying volatility. Coordinate communications if your company spans regions.
Finally, pair news consumption with internal drills. Tabletop exercises for agent security, treasury policy, or SEO measurement catch gaps that headlines merely describe.
Comments
Loading comments…