[ new ]Claude Managed Agents on Superserve
[ About · Est. 2024 · San Francisco ]

We're building the runtime agents were missing.

Superserve orchestrates millions of microVM sandboxes. Fast enough to spin up per LLM call, durable enough to keep an agent's filesystem alive for weeks, and isolated enough that the code inside them doesn't matter.

[01]12B+Sandboxes run
[02]99.97%Trailing 90d uptime
[03]22Humans, shipping
[04]$8MSeed · Felicis lead
[01] ManifestWhy we exist

The infrastructure agents were never given.

manifest.ts
// 2023. The agent stack started looking like this:const stack = {  reasoning: "miraculous",  tooling:   "shipping weekly",  runtime:   "a 15-minute Lambda or a leaky container",};// Models could write code, plan, browse, debug. But the place they ran in// timed out, lost state, or shared a kernel with someone else's sandbox.// We built Superserve so that the runtime stops being// the part of the stack that breaks first.+ Sub-300ms cold starts, on a per-call microVM.+ Real virtual filesystems. sqlite, pnpm, ffmpeg, 5GB datasets.+ Snapshots that survive process death, region failure, and 30 days idle.+ Hardware isolation, every call. The model can't escape.// And the same thing, in your own cloud, on your own bill, if you need it.
[02] Operating principlesHow we work

Four things we don't compromise on.

01

Reliability is the feature.

Agents that disappear at 24h, time out at 15m, or quietly drop file writes are not actually working. We optimize for the boring metrics: p99 latency, sandbox survival rate, snapshot durability.

02

Security is not a checkbox.

Every sandbox is a hardware microVM. No shared kernels, no "trust the user" escape hatches. We assume every line of code we run is hostile, because for our customers it usually is.

03

Open by default.

Our SDK, CLI, and benchmarks are MIT-licensed. Our control plane runs in your cloud if you want it to. Lock-in is a tax on customers, not a moat.

04

Talk to the product, then to humans.

You should be able to ship to production without a sales call. When you do need it (incident at 3am, custom region, weird workload), a human picks up.

[03] CustomersIn production

Trusted by teams running real agent workloads.

Kernel.ai
Tinsel
Roam Labs
Forgehaus
Atlascose

"We replaced our in-house container pool with Superserve in a weekend. Cold starts went from 4 seconds to under 300ms, and we stopped getting paged when long-running agents hit the 24-hour wall. It's the boring kind of infrastructure win. Everything just works."

Priya Shah · Head of agents, Kernel.ai
[04] Security postureBy default

What we do, not what we promise.

Compliance

SOC 2 Type II · HIPAA

Audited annually by Prescient. BAA available on Pro and Self-hosted. Pen-test report on request.

Isolation

Hardware microVMs

Every sandbox is a Firecracker microVM with its own kernel. No shared anything. We treat the workload as hostile.

Data

Encrypted, segregated

TLS in transit, AES-256 at rest, per-project KMS keys. BYOK on Self-hosted. Sandbox state never leaves region.

Access

SSO + RBAC

SAML, OIDC, SCIM. Fine-grained roles for orgs, projects, sandboxes. Per-key network egress allow-lists.

Observability

Full audit trail

Every exec, file write, network egress, logged, queryable, exportable to your SIEM. Retention up to 7 years.

Uptime

99.97% trailing 90d

Public status page at status.superserve.ai. SLA up to 99.99% on Self-hosted. Multi-region, automatic failover.

[05] TimelineHow we got here

A short history.

Q1 · 2024

Founding

Three of us (ex-Anthropic infra and two from a Firecracker-shop) got tired of watching agent demos die at the 24-hour mark. Started prototyping in a basement in SF.

Q3 · 2024

Private beta

First 12 design partners. Ran 4M sandboxes in three months across coding agents, RL training, and a stealth biotech doing simulation rollouts.

Q1 · 2025

Seed round

$8M from Felicis, Conviction, and a handful of agent-platform founders. Hired the first 6 engineers. Shipped snapshots & forks.

Q3 · 2025

GA + open source

Public launch. Open-sourced the SDK, CLI, and our benchmark suite. SOC 2 Type II audit complete.

Q1 · 2026

Self-hosted GA

BYOC general availability. Three Fortune-500 banks and one US-Federal agency now run Superserve inside their own VPC.

[06] Team22 people · SF + remote

Built by people who've broken this before.

AR
Ada Reyes
CEO · Co-founder
ex-Anthropic infra · Stripe
MS
Mikkel Sørensen
CTO · Co-founder
ex-Cloudflare Workers · Firecracker
PI
Priya Iyer
Head of Engineering
ex-Vercel · GitHub
JW
Jonas Weber
Co-founder
ex-AWS Lambda · Manuva
SL
Sara Lin
Security Lead
ex-Outsider NSA TAO
RM
Ravi Menon
Platform Engineer
ex-Modal · Replit
NA
Nora Achebe
Developer Experience
ex-Vercel · Render
HT
Hiro Tanaka
GTM
ex-Datadog · Snowflake
[07] Investors$8M seed · 2025

Backed by people who have shipped infra.

Felicis
Conviction
Y Combinator
Base Case
South Park Commons
Scale Angels

Plus angel investments from founders of Vercel, Cloudflare, Modal, Stripe, and Replit.

[08] CareersWe're hiring

Help us build the runtime agents deserve.

We're a small team, in-person three days a week in SF, remote-friendly for the rest. We hire seniors, pay above market, and ship to production every day.

Staff Systems EngineerSF · onsite hybrid
Security Engineer (microVM)SF or remote
Developer Relations LeadSF or remote
Forward Deployed EngineerSF · onsite
See all roles ↗
[ The offer ]

What you get.

  • Above-market cash + meaningful equity
  • Health, dental, vision (US) · stipend (intl)
  • 401(k) with 4% match · 16 weeks parental
  • $3k/yr learning budget, $2k/yr setup
  • Real PTO. 25 days plus all the holidays
  • Quarterly team off-sites

Try it before you talk to anyone.

npm install, paste a key, ship a sandbox. The sales call is optional and we'll never make it required.