Launch HN: Rivet (YC W23) – Scale and Orchestrate Agents on S3
TL;DR Highlight
Running each agent as an Actor gave 4.99 ms p99 cold starts and 72 KB RAM per instance.
Who Should Read
Backend developers whose per-agent sandbox VMs are costly and slow to start.
Core Mechanics
- One agent equals one Actor
- An Actor is a long-lived process with in-memory state, SQLite, WebSockets, and scheduling.
- You can create one per agent, user, session, or tenant.
- Numbers vs sandboxes
- An Actor shows 4.99 ms p99 cold start and 72.4 KB memory, versus 1.70 s and 512 MiB for a sandbox (E2B).
- Actors share a node process, so one server holds hundreds of thousands.
- S3 is the only dependency
- All data lives in S3, with no Postgres, FoundationDB, or NATS needed.
- GCS, Azure Blob, MinIO, and R2 also work.
- Agent loop runs outside the sandbox
- The premise is that agents mostly wait on inference while a VM sits idle.
- State, memory, and context stay in the Actor, which reaches the sandbox through tools only to run code.
- You may not need a sandbox
- Agents that only use tool calls, MCP, and APIs run without one.
- For code execution, use agentOS (V8/WebAssembly) or attach your existing sandbox as tools.
- Pitched as a Durable Objects alternative
- It says it has no 128 MB memory cap and no 10 GB SQLite limit.
- It runs on Node.js or Bun without a custom runtime and is Apache 2.0.
Community Reactions
- Storage model explanation landed well
- Commenters found keeping S3 off the write path (2 of 3 AZ SSDs, then batched to S3) easy to understand.
- One opinion said a POSIX filesystem interoperating with S3 would be interesting.
- Real-world usage shared
- One user moved a Cloudflare Durable Object workload into their own VPC with Rivet and found it pleasant.
- Another reported running Rivet v3 self-hosted on AWS.
- Complaints that it's unclear what it does
- One opinion said the site should lead with the problem rather than benchmarks.
- Others found the actor vs agent distinction unclear and the UX confusing, even for coding agents.
- Pushback on performance claims
- One opinion said orchestration speed isn't an agent bottleneck since most time is spent waiting on the model.
- Others asked for numbers backing the claim that sandboxes are expensive and slow, and for a security comparison with E2B.
How to Apply
- Measure wait time in your agent loop
- Log model wait, tool execution, and idle sandbox time for one session today.
- If the sandbox is mostly idle, the loop is a candidate to move out.
- Split agents by sandbox need
- Pick agents that only use tool calls, MCP, and APIs, like customer support.
- Try running them in an Actor example without a sandbox.
- Store state in per-session SQLite
- Restructure to save history and artifacts in a per-agent SQLite table.
- Verify locally that history survives a restart.
- Connect code execution only as a tool
- Keep your sandbox provider and wrap it as a bash tool the Actor calls.
- To shrink the trust boundary, move just code execution to a lightweight sandbox like agentOS.
Try It
# Paste into Claude Code, Cursor, Codex, or any coding agent (the README's own prompt)
Run npx skills add rivet-dev/skills, then ask me what to build and run it locally.
# Answer the agent's questions with {{the agent you want to build}} and confirm it runs locally
# Next: restart the agent and check that its conversation history persistsTerminology
Related Papers
Claude Code sends your email address to websites
Claude Code put users' emails in curl User-Agent headers, prompting spam complaints and a built-in blocking mod
Show HN: Stroq – a local audit of what your coding agent did and why
Stroq finds leaked keys in agent sessions and traces risky commands to their source before they run
Building a Multi-Agent System from Scratch: Architecture Patterns
Splitting work across agents adds coordination problems, so enforce hard limits in code.
On Turn-Level Best-of-N Sampling for Scaling LLM Agents
Picking one of N candidate actions per turn with an LLM raised pass@1 and cut agent turns by 40–50%.
When Do Summaries Become Instructions? Unsupported Policy in LLM Agents' Self-Written Handoff Summaries
Making the system prompt's approval sentence specific cut unsupported restrictions in agent summaries from 5.0% to 1.8%.
AI is forcing us to pay off 40 years of technical debt
AI turned an old PHP crash bug into a code-execution exploit within two weeks.
Postman silently auto-installs AI plugins and attempts to share all API data
Opening Postman auto-installed plugins and a startup hook into Codex and Claude Code
Octomind – open-source AI harness that saves your tokens, static rust
A single Rust binary runs any model as a coding agent, with TOML guardrails replacing approval prompts.