🔥 Trending on HN

Rivet’s bet on sleeping AI agents and S3 storage

3 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Rivet Actor(REE-vit AK-ter)

A long-running process that keeps an agent’s work and data.

S3(ess three)

A service for storing large amounts of data.

p99(pee ninety-nine)

A test point that covers 99 percent of measured cases.

What happened

Rivet is an open-source platform for running and coordinating AI agents. Its Launch HN post presents Rivet Actors as long-lived processes. Each Actor can hold memory, a SQLite database, WebSockets, and scheduled work. It can represent an agent, user, session, or tenant. It can also hibernate when idle and wake with its data later. The post drew 120 points and 59 comments on Hacker News, a technology discussion site. Those figures show community attention. They do not prove Rivet’s claims are correct.

Why this design exists

Rivet’s explanation starts with waiting. An agent often spends much of its time waiting for model inference. It may need a sandbox only when it runs code. A full sandbox can remain mostly idle during those waits. Rivet moves the agent’s state, memory, and context into an Actor. The Actor reaches a sandbox through tools only when needed. Rivet describes this as separating the agent’s head from its hands.

How Rivet works

Rivet’s architecture guide divides the system into three parts. Object storage keeps Actor data. A control plane schedules Actors and routes requests. Nodes run the application code. Each Actor has a SQLite database. Recent data stays on SSDs. Older data is batched into S3 or another compatible object store. A sleeping Actor can release memory and wake on another node with its saved data.

This design gives each agent a durable place to live. Rivet lists durable agents, workflows, AI-generated apps, and collaborative editing as use cases. The project is open source under the Apache 2.0 license. It can be self-hosted. Its documentation says an object-storage bucket can be the main external dependency.

Why it matters

Many AI products need more than a model response. They need session history, files, schedules, live connections, and recovery after restarts. Rivet tries to provide those pieces in one Actor. That could make large groups of agents easier to keep available without giving every agent a separate virtual machine.

The important point is the architecture, not a promise that every workload becomes cheap. The project supports cloud deployment, self-hosting, and several object-storage providers. Operators would still need to choose how much isolation, storage speed, and code execution safety each workload requires.

What the reported numbers mean

Rivet’s official benchmark reports a 4.99 millisecond p99 cold start and 72.4 KB of resident memory per running Actor. It compares those figures with an E2B sandbox at 1.70 seconds and 512 MiB. Rivet also notes that the measurements use different environments and methods. The comparison is therefore useful as context, not as a universal speed ratio.

Rivet also presents storage savings of about 10.4 times against replicated SSD storage. That figure comes from AWS list-price calculations. It is not a measured bill. The calculation excludes several costs, including some transfer, performance, and local-disk costs.

What remains unknown

The sources do not show how Rivet performs across many real production workloads. They do not establish the total cost of frequent reads, writes, wakeups, or failures. They also do not show how every deployment should isolate untrusted code. Rivet’s own guide still recommends using a sandbox when an agent must run code it wrote.

What to watch next

The next useful evidence would be independent benchmarks, reproducible tests, customer deployments, and detailed cost reports. It will also matter how often Actors sleep and wake, how quickly cold data returns from object storage, and how operators combine Actors with sandboxes. Hacker News attention makes Rivet worth examining. It is not a substitute for that evidence.

💬 Rivet is an actor-based foundation for keeping agent state lightweight and durable

The comments focus on using SQLite and SSDs as a fast tier, S3 as a storage tier, and separating the agent loop from the sandbox. They also leave open questions about the evidence for sandbox-speed claims, cross-AZ cost, wake-up safety, and how clearly the product is explained.

  • Rivet is presented as an actor-model foundation built around long-lived, lightweight actors with state. Several comments treat it as a layer for hosting agent loops and tool calls, rather than treating actor and agent as interchangeable terms.
  • The developer says Claude Code and Codex can run inside agentOS, but describes the preferred architecture as keeping the agent loop outside the sandbox and exposing the sandbox as a tool. The stated reasons are a smaller trust boundary and blast radius, plus avoiding the memory cost of a heavyweight process for every agent.
  • The storage description in the comments says an actor's SQLite write is acknowledged after SSDs in 2 of 3 availability zones have it, then batched to S3 in the background. Reads use the SSD cache when available and fall back to S3 otherwise, with tunable cache capacity. This is intended to keep S3 off the write path and reduce latency and cost, but the thread does not provide an independent benchmark of it.
  • One user's self-reported estimate warned that the cross-AZ traffic implied by this design could cost at least $0.04 per GB on AWS. The real bill depends on region and configuration, and the estimate is not verified in the thread.
  • The strongest performance objection is that, when model inference dominates the wait, sandbox cold starts may not be the main bottleneck. A commenter asked for numbers behind the claim that sandboxes are expensive, mostly idle, and slow to start. The developer's response was that workloads differ and that Rivet complements rather than eliminates sandboxes.
  • There is also a product-clarity objection. A technically experienced commenter said the description did not make the actor-versus-agent distinction or the solved problem clear, and requested use-case tutorials before benchmark-focused messaging. That is user feedback about comprehension, not proof that the system is slow.
  • Hibernate and resume raise durability questions. One user reported that if tool results remain only in memory or a temporary file instead of the same durable snapshot, a resumed agent may invent a confident continuation. Other commenters asked how to fence split-brain during a network partition while an S3 batch is draining, and whether a cited 138 ms cold read becomes slower for larger transcripts and filesystems. These are reported failure modes and open questions, not confirmed Rivet bugs.

mature digest at 59 comments (revision 1). We fetched 59 comments and sampled 59 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Rivet lets many AI helpers sleep between tasks

📰 Full story: Rivet’s bet on sleeping AI agents and S3 storage

Rivet is a platform for running many AI agents. Its main idea is simple: keep their work safe, then let idle agents sleep.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
AI agent(ay eye AY-jent)

A computer helper that follows instructions and does tasks.

Actor(AK-ter)

A lasting space that keeps one helper’s work.

S3(ess three)

A service that stores data.

💡 The gist

  • Rivet runs many AI agents in small, lasting spaces.
  • Idle agents can sleep and wake with their saved work.
  • Hacker News gave the post 120 points and 59 comments. That shows attention, not proof.

Rivet is an open-source platform for running AI agents. An AI agent is a computer helper that follows instructions and completes tasks.

Rivet gives each agent an Actor. An Actor is a long-running process with memory, saved records, and connections. It can represent one agent, user, session, or group. When nobody needs it, the Actor can sleep. When a new request arrives, it can wake and continue.

Why does this help? An AI agent often waits for a model to answer. It does not run code every second. A full sandbox can sit mostly unused during those waits. Rivet keeps the agent’s state outside the sandbox. The agent uses a sandbox only when it needs to run code.

Rivet stores recent information on fast SSDs. It moves older information to S3, an object-storage service. This gives each Actor a lasting place for session history and other files. Rivet says it can run in the cloud or on a company’s own machines. It uses the Apache 2.0 open-source license.

Rivet’s official tests report very fast starts and low memory use. They compare an Actor with an E2B sandbox. However, Rivet says the tests used different setups. That means the results give useful context. They do not guarantee the same speed everywhere. The storage savings are also a price calculation, not a bill from a real customer.

The big question is practical. Can Rivet stay reliable and affordable with real users, frequent wakeups, large files, and failures? Its benchmark notes explain the setup. Hacker News interest is a reason to look closer, not a final answer.

💬 Rivet is a way to let AI work, sleep, and keep its memory

Rivet keeps an agent's work in small actors, using fast SSDs and larger S3 storage together. The idea is promising, but users still question its cost, explanation, and safety when sleeping agents wake up.

  • Rivet keeps state in a small, long-lived actor. The actor can hold an agent loop and its tools; actor and agent are related but not the same thing.
  • The comment's storage explanation says data is accepted after reaching SSDs in 2 of 3 zones, then copied to S3 in batches. Reads come from SSD when possible and S3 otherwise. One user self-reported an estimate of at least $0.04 per GB for cross-zone traffic, but the actual cost was not verified.
  • The agent's thinking loop can stay outside a sandbox while the sandbox acts like a tool. The developer says this can make the trust boundary and memory use easier to manage, even though running the agent inside agentOS is possible.
  • Some users argue that model waiting time may matter more than sandbox startup time. Others say the product is hard to understand and needs clearer examples. The developer says the right choice depends on the workload and that sandboxes still have a role.
  • A sleeping agent must save tool results as well as conversation state. Otherwise it might wake up and make up the next step. Users also ask whether the cited 138 ms read gets slower with bigger files and whether a network break could create conflicting copies. These are questions and reported risks, not confirmed Rivet bugs.

mature digest at 59 comments (revision 1). We fetched 59 comments and sampled 59 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Rivet helps computer helpers sleep

📰 Full story: Rivet’s bet on sleeping AI agents and S3 storage

Rivet is a tool for computer helpers. It lets them rest when nobody needs them.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Rivet(REE-vit)

A tool that runs computer helpers.

S3(ess three)

A big place for saving data.

Hacker News(HACK-er news)

A website where technology readers discuss projects.

Rivet is a tool for running AI helpers.

An AI helper is a computer that does jobs.

Rivet gives each helper a little work space.

The space is like a small room.

It is not a real room.

The helper keeps its work there.

When nobody needs it, the helper sleeps.

Its notes go into S3, a big storage service.

The helper wakes and reads those notes.

Then it can continue its job.

Hacker News gave Rivet 120 points and 59 comments.

That means many readers noticed it.

It does not prove Rivet works perfectly.

Rivet’s own tests say it can start quickly.

Other computers may get different results.

You can read Rivet’s project page.

💬 Rivet helps little AI workers remember their jobs

Rivet gives AI workers a place to keep their work while they sleep. It uses a fast shelf and a big storage room, but people still want to test its cost and safety.

  • Rivet lets a small AI worker keep its memory and sleep until it is needed again. The worker is called an actor; an actor and an agent are not exactly the same.
  • The design described by a commenter puts new memory on SSDs in 2 of 3 places, then sends it to an S3 storage room. One user's estimate says the trip between places might cost at least $0.04 per GB.
  • The AI's thinking part can stay separate from its little tool box. People disagree about whether the tool box is really the slowest part, and some say Rivet needs clearer examples.
  • If the worker forgets what a tool told it, it might confidently do the wrong next thing. Users also ask about the cited 138 ms after a long sleep and what happens when the network breaks.

mature digest at 59 comments (revision 1). We fetched 59 comments and sampled 59 across the thread. These are HN users’ reports, not independently verified facts.

Sources