Rivet's Launch HN Claims 72 KB Per Agent and 4.99 ms Starts, With No Workload Named

Rivet, an open-source orchestrator for AI agents, claims each agent can run as a 72 KB "Actor" that starts in 4.99 milliseconds at the 99th percentile, against 512 MB and 1.7 seconds for a minimal sandbox. Founder Nathan Flurry posted the numbers in a Launch HN thread on Nov. 5, 2026, and they are the company's own. Neither the README nor the post names the workload, hardware or dataset behind them.

The thread, titled "Launch HN: Rivet (YC W23)", reached 120 points and 59 comments. The code is Apache 2.0 on GitHub, where the repository shows about 6.8 thousand stars, 312 forks, 48 open issues and 6,324 commits.

What Rivet sells

An Actor, in the README's words, is a "long-lived process with in-memory state, a SQLite database, WebSockets, and scheduling". It hibernates when idle, wakes on the next request and stores data in S3. Writes are acknowledged once SSDs in two of three availability zones hold them, then batched to S3. The post gives p99 latencies of 1.07 milliseconds for durable writes, 1.36 milliseconds for reads and 138 milliseconds for reads from object storage, and says storage costs about 10 times less than EBS gp3.

There are three ways to run it: a single Rust binary plus S3 for self-hosting, Rivet Cloud on dedicated AWS or GCP clusters, and a bring-your-own-cloud tier where Rivet manages a control plane in the customer's VPC. The README lists no prices for any tier. Python and Swift clients are marked experimental and Dynamic Apps are in preview.

Metric (founders' figures)	Rivet Actor	Minimal sandbox
Memory baseline	72 KB	512 MB
p99 cold start	4.99 ms	1.7 s
p99 durable write	1.07 ms	not given

AI Weekly, which wrote the launch up on Nov. 6, adds comparisons against Cloudflare Durable Objects: 39 times faster actor creation, 28 times faster SQLite writes and 13 times faster backend requests, all attributed to Flurry. It also reports that agent startup Amp moved from Durable Objects to Rivet over a weekend, a claim that also appears in the launch post.

The batch is listed two ways

Hacker News carries the title "YC W23", and Flurry's post names co-founder Nicholas, with no surname. AI Weekly calls Rivet a YC F25 company. The Terminal could not reconcile the two from the pages it fetched. Treat the batch as unconfirmed until Y Combinator's company directory settles it, because a 2023 company launching a product in 2026 and a 2025 company doing so are different stories about funding and runway. No funding amount appears in any source.

The sources also differ on a smaller number. The GitHub page shows about 6.8 thousand stars, while AI Weekly reports 6.7 thousand at launch, probably because they were read at different times.

What commenters asked

The criticism was specific. User esafak argued that orchestration speed is not the bottleneck because model latency dominates, and suggested leading with durability. User chicagobuss warned that cross-zone replication could add roughly $0.04 per GB in AWS charges. User takeitwhen raised the risk that if tool results are not in the durable snapshot, a woken actor may invent a continuation. User ATechGuy2 asked for data behind the claim that sandboxes are expensive and slow.

Flurry replied that the sandbox claims depend on workload, and that some Fortune 500 companies use Rivet to cut costs on reinforcement-learning workloads. He did not name them. The pattern matches Momentic's benchmark built on its own bug set and Billion Context's results in the author's own repo: a clean ratio from the vendor, with the method left to the reader.

What to check next

The README's comparison pages are the first place to look for a method. Until someone outside Rivet publishes a run against a named workload, 72 KB is a design target, and the Amp migration is the only customer detail on record.