Private beta · Q1 2026

Agents that
survive the night.

Cinder is a durable execution layer for long-running AI agents. Every step is checkpointed, every failure is recoverable, and nothing is lost when the process dies at 3 a.m.

No spam. One email when the beta opens.

Waitlist
2,418
Design partners
14
Steps replayed
9M+

01 What we're building

Three primitives.
Nothing else.

We are not shipping a framework, a dashboard, or a dozen integrations. Cinder is deliberately small: a journal, a scheduler, and a replay engine.

01

The journal

Every tool call, token, and side effect is appended to an immutable log. Steps are idempotent by construction, so a retry is never a duplicate charge.

  • Append-only, Postgres-backed
  • Per-step idempotency keys
  • Point-in-time inspection
02

The scheduler

Long horizons, short budgets. Cinder keeps agents alive across hours, days, and deploys without a cron job duct-taped to a queue.

  • Durable timers up to 90 days
  • Fair-share concurrency
  • Backpressure that doesn't lie
03

The replay engine

Reproduce any run, step by step, with the exact model output it saw. Debugging becomes reading, not guessing.

  • Deterministic re-execution
  • Fork a run from any step
  • Diff two trajectories

02 In practice

Write it like
it can't fail.

No state machine, no queue plumbing, no retry ladder. You describe the work; Cinder decides what to persist and when to resume.

  • 01 Steps are memoized by key
  • 02 Failures resume, not restart
  • 03 Runs outlive the process
research-agent.ts
1import { cinder } from "@cinder/sdk";
2
3const agent = cinder.agent("research", {
4  model: "gpt-5",
5  budget: { steps: 400, wall: "6h" },
6});
7
8await agent.run(async (step) => {
9  const plan = await step.call("planner", () =>
10    planTask(input)
11  );
12
13  for (const task of plan.tasks) {
14    await step.call(`fetch:${task.id}`, () =>
15      fetchSource(task), { retries: 3 }
16    );
17  }
18
19  return step.checkpoint("draft", () =>
20    writeDraft(plan)
21  );
22});
resume point If the process dies at line 15, Cinder replays from the last checkpoint — not from zero.

03 Where we are

Honest status,
updated weekly.

We publish the build log because a beta is a promise, not a launch. Here is exactly what works today and what is still on the bench.

Journal & replay core92%
Scheduler74%
Console38%
Self-hosted installer21%
Most agent failures aren't model failures. They're process failures — a dropped connection, a redeploy, a timeout nobody wrote down. We're building the part that remembers.
Ada Kessler co-founder, engineering
i.

Boring on purpose

Postgres, object storage, and a queue. No new database to learn at 2 a.m.

ii.

Your data stays yours

Self-hosted from day one. The cloud is a convenience, never a requirement.

iii.

No dashboards for vanity

If a metric doesn't change a decision, it doesn't ship.

04 Early access

Be there when
it opens.

Fourteen design partners are already running Cinder in staging. The next cohort opens in Q1 2026 — waitlist members get in first.

We'll only write when there's something to run.