Durable runtime for AI agents · v4.2

AI workflows thatshow their work.

Orrindel runs your agents and automations as durable steps. Every call is retried, every decision is logged, and any run can be replayed from the step that broke.

Start building →

Steps run today

48,123,379

Retries absorbed this week

212,904

p99 step overhead

0 ms

Runs lost to crashes

Runs in production at

◆Halvard Health▲NORTHPIER●kestrel.ai■Tallow & Co◐Quanta Freight✦meridel▮BRIGHTWELL◇sable/ops◆Oxbow Legal▲Pellam

01 · The problem

Most agent code forgets everything when it crashes.

A script holds its progress in memory. When the connection drops on step five, the restart runs steps one to four again, and your customer gets refunded twice.

Steps redone

Duplicate refunds

Time lost

refund-agent.js · pm2 logsidle
14:02.118$ node refund-agent.js --ticket 48213
14:02.402[1/6] classify ............ ok  intent=refund (1.21s)
14:03.611[2/6] fetch account ....... ok  ac_7Q2 (0.33s)
14:03.944[3/6] find charges ........ ok  ch_31x duplicate (0.54s)
14:04.487[4/6] issue refund ........ ok  re_88k $49.00 (0.27s)
14:04.760[5/6] draft reply (llm) ... …
14:41.003Error: read ECONNRESET at TLSSocket.onStreamRead
14:41.004process exited with code 1
14:46.220pm2: restarting refund-agent (attempt 2)
14:46.531[1/6] classify ............ ok  (again · 1.19s · $0.0009)
14:47.722[2/6] fetch account ....... ok  (again)
14:48.061[3/6] find charges ........ ok  (again)
14:48.603[4/6] issue refund ........ ok  re_91c $49.00  ← second refund
# the customer now has two refunds and one confused support agent.

02 · Durable execution

When a step breaks, the run doesn’t start over.

Orrindel saves the result of every step. Fix the bug, press replay, and the run picks up at the step that failed, with everything before it intact.

01RunFive steps start. Each result is saved as it lands.
02Failissue-refund sends no amount. Retries can’t fix a bug, so it stops at 3/3.
03FixChange one token. Deploy. The failed run is still waiting.
04ReplaySteps 1–3 come back from memo in 0 ms. Only the broken step runs again.

Steps restored

Time not re-spent

0.00s

LLM calls repeated

Duplicate refunds

refund-flow.tsrun_01JAX9TQticket #48213● running↺ Replay from step 4
1export const refundFlow = orrindel.workflow(
2  { id: "refund-flow", trigger: "ticket.created" },
3  async ({ event, step }) => {
4
5    const intent  = await step.run("classify", () => llm.classify(event.body))
6    const account = await step.run("fetch-account", () => crm.get(event.customerId))
7    const charges = await step.run("find-charges", () => billing.list(account.id))
8    const refund  = await step.run("issue-refund", () =>
9      billing.refund(charges.duplicate.id, charges.duplicate.amount))
10    await step.run("reply", () => mail.send(event.from, refunded(refund)))
11  }
12)
 StepTimeSaved output
●classify0.00s
·fetch-accountqueued
·find-chargesqueued
·issue-refundqueued
·replyqueued
attempt log
replay—
0s1s2s3s4s

03 · Primitives

Six calls cover almost every workflow you’ll write.

No new language and no YAML. You write normal functions and wrap the parts that talk to the outside world.

step.run()01

Run it once, keep the result

Wrap any call. The result is saved, so a retry or replay reads it back instead of calling again.

const acct = await step.run("fetch-account",
  () => crm.get(event.customerId))

step.sleep()02

Wait days without a server

Pause a run for minutes or months. Nothing polls and nothing is billed while it sleeps.

await step.sleep("trial-ends", "14d")
await step.run("send-offer", sendOffer)

step.waitForEvent()03

Pause until something happens

Hold a run until a matching event arrives, like a paid invoice or a signed form, with a timeout.

const paid = await step.waitForEvent("paid",
  { event: "invoice.paid", timeout: "7d" })

retries04

Retry with backoff, by default

Failed steps retry on their own: 3 times, with exponential backoff. Errors you mark final stop at once.

step.run("charge", charge, {
  retries: 3, backoff: "exponential" })

concurrency05

Limit how much runs at once

Cap parallel runs per key, like one per customer, so a burst from one account never starves the rest.

orrindel.workflow({ id: "sync",
  concurrency: { limit: 2, key: "event.accountId" } })

step.invoke()06

Call another workflow

Compose workflows. The parent waits for the child's result, and both show up in one trace.

const lead = await step.invoke("enrich",
  { function: enrichLead, data: { email } })

04 · Human in the loop

Let a person say yes before the agent does anything permanent.

Any step can pause for a decision. The run waits, for seconds or for days, then carries on down the branch the person picked. Every answer is recorded with who gave it and when.

Timeout: 24h, then escalates to #support-leads
Channels: Slack, Teams, email or your own UI
Audit: who, when, and what they saw
Cost while waiting: $0.00

orrindel#support-approvals09:22

Refund needs approval: $49.00 to Jamie Alvarez for ticket #48213. The March charge appears twice.

amount
$49.00 USD
charges
ch_31x · ch_30q (Mar 3)
agent confidence
0.94
policy
refund window 60d ✓

◷ expires in 23:59:59

run_01JAX9TQrefund-flow
✓draft_reply1.60s · 182 words
◷wait_for_approvalwaiting · 0s
·send_refundqueued
·close_ticketqueued

05 · Flow control

One noisy customer shouldn’t slow down everyone else.

Set a concurrency limit and a fairness key, and Orrindel serves each account in turn. A burst from your biggest customer no longer puts the smallest one at the back of the line.

concurrency 4
acct_bigco · 40 jobs
0/40 done · waited p95 0.0s
acct_ridge · 8 jobs
0/8 done · waited p95 0.0s
acct_tern · 3 jobs
0/3 done · waited p95 0.0s
t = 0.0srunning 1/4 queued running doneconcurrency: { limit: 4, key: "event.accountId" }

06 · Observability

See every prompt, tool call and token, in the order it happened.

Each run keeps a full trace: what the model was asked, what it answered, which tools it called and what that cost. Hover a span to open it.

span0s2s4s6sms
support-triage6800
└ classify1200
└ search_docs800
└ embed_query170
└ fetch_account420
└ draft_reply3200
└ lookup_export340
└ requeue_export220
└ grade_reply700
└ send_reply500
LLM calltool calleval
draft_replyLLM · 3.20s
model
large
tokens
2,904 → 312
cost
$0.0031
tool calls
2
INPUT
System: You are a support engineer… Context: 4 passages, account ac_2Hv (business, eu-west)…
OUTPUT
"Hi Priya, large exports over 2 GB were timing out in eu-west this morning. We’ve raised the limit and re-queued yours…"

spend · last 24h$50.42 total1.9M tokenspeak 14:00 · $4.60budget alert at $80/day · ok
00:0006:0012:0018:0023:00

07 · Tools

Every integration is a typed function your agent can call.

Connect CRMs, billing, mail, chat, databases and browsers as typed tools. The types become the schema the model sees, so a bad tool call fails before it runs, not after.

crm· cached 60s
crm.get(customerId: string)
→ Promise<Account>

billing· idempotent
billing.refund(chargeId: string, amount: Money)
→ Promise<Refund>

mail· retries 3
mail.send(to: Email, body: Markdown)
→ Promise<MessageId>

chat· rate 1/s
chat.post(channel: string, text: string)
→ Promise<Message>

tickets· idempotent
tickets.update(id: number, patch: TicketPatch)
→ Promise<Ticket>

vector· p95 180ms
vector.search(query: string, k?: number)
→ Promise<Passage[]>

sheets· retries 5
sheets.append(sheetId: string, row: Cell[])
→ Promise<RowRef>

calendar◷ needs approval
calendar.book(slot: TimeRange, who: Email[])
→ Promise<Event>

payments· idempotent
payments.capture(intentId: string)
→ Promise<Payment>

docs· retries 3
docs.create(title: string, body: Markdown)
→ Promise<DocRef>

sql· read-only role
sql.query(text: SQL, params: unknown[])
→ Promise<Row[]>

browser· timeout 90s
browser.open(url: URL, actions: Action[])
→ Promise<PageSnapshot>

pdf· retries 2
pdf.extract(file: FileRef, schema: Zod)
→ Promise<T>

http· circuit breaker
http.request(req: RequestInit & { url })
→ Promise<Response>

storage· idempotent
storage.put(key: string, data: Blob)
→ Promise<ObjectRef>

sms◷ needs approval
sms.send(to: Phone, text: string)
→ Promise<SmsId>

queue· exactly once
queue.publish(topic: string, msg: Json)
→ Promise<Offset>

llm· tokens metered
llm.complete(prompt: Prompt, tools?: Tool[])
→ Promise<Completion>

embeddings· batched 96
embeddings.create(input: string[])
→ Promise<Vector[]>

webhooks· deduped 24h
webhooks.receive(source: string, verify: Secret)
→ Promise<Event>

120+ ready-made tools·bring your own with one function·MCP servers supported

08 · Your language

Write it in the language you already ship.

Workflows live in your repo, deploy with your app and run on your cloud or ours. TypeScript, Python and Go SDKs share one runtime and one dashboard.

workflows/onboarding.ts
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

STEP 01
Install the SDK
$ npm i @orrindel/sdk
or pip install orrindel · go get orrindel.dev/sdk
STEP 02
Run the dev server
$ npx orrindel dev
Local dashboard on :8288, with replay and traces
STEP 03
Deploy with your app
$ git push
Workflows register on boot. No separate workers.

09 · Use cases

Start from a workflow someone already shipped.

Eight production patterns with the steps, the tools and the numbers. Each one is a template you can fork into your repo.

All use cases →

Support

Refunds that approve themselves, mostly

Classify refund tickets, check the charge history, refund duplicates under $200 automatically and ask a lead before anything bigger.

7 steps

ticket.created

Sales

Enrich and route every inbound lead in 40 seconds

Look up the company, score fit with an LLM against your ICP, write the first reply and assign an owner, before the visitor has closed the tab.

6 steps

form.submitted

Data

Turn a folder of contracts into a searchable index

Extract text and clauses from every uploaded PDF, chunk and embed them, and keep the index in sync when files change or disappear.

9 steps

storage.object_created

Finance

Invoices in, ledger entries out, with a human on exceptions

Read invoices from a shared inbox, match them to purchase orders, post clean ones to the ledger and send mismatches to finance with the reason attached.

8 steps

mail.received

10 · From the teams

What engineers wrote in the commit that switched them over.

$ git log --author-date-order --grep=orrindel
a41f9c2chore: delete our home-grown queue and retry layerMaya Okafor · 3 days ago
MO“We moved claims intake onto Orrindel in two sprints. The pager for stuck jobs used to go off weekly. It hasn’t gone off since March.”
Maya Okafor · Staff engineer, Halvard Health
+212−2,318++----------------------
7c02be8feat: replay failed shipments from the failing stepTomás Reyes · 2 weeks ago
TR“Carrier APIs fail all the time. Now a failed label run waits, we fix the mapping, and replay picks up at step four. Nobody re-keys an address.”
Tomás Reyes · Head of platform, NORTHPIER
+96−431++++--------------------
e93d114perf: memoize the expensive eval stepsHana Sato · 1 month ago
HS“Our eval pipeline makes 14 model calls per document. When step 12 flaked we used to pay for all 14 again. Memoized steps cut our eval bill by 38%.”
Hana Sato · ML engineer, kestrel.ai
+58−190++++++------------------
1b7a6f0feat: human approval before any refund over $200Jonah Whitfield · 2 months ago
JW“Legal wanted a person in the loop for refunds. It took one step.waitForEvent and a Slack button. Audit asked for the log and we just exported it.”
Jonah Whitfield · CTO, Tallow & Co
+44−12+++++++++++++++++++-----
5fd20acfix: stop one tenant starving the othersPriya Raman · 3 months ago
PR“A single customer’s nightly import used to block everyone’s morning syncs. A concurrency key per account fixed it in one line.”
Priya Raman · Senior SRE, Quanta Freight
+3−1++++++++++++++++++------

11 · Security

Built for the teams your security review will send.

Your data stays in the region you pick, encrypted with keys per workspace. Enterprise can run the whole runtime inside your own cloud account.

Compliance

SOC 2 Type II, GDPR, HIPAA with a signed BAA

Data residency

us-east, eu-west or ap-south, pinned per workspace

Encryption

AES-256 at rest with per-workspace keys, TLS 1.3 in transit

Access

SAML SSO, SCIM provisioning, scoped API keys

Audit

Every run, approval and config change, exportable to your SIEM

Self-hosting

Run in your AWS, GCP or Azure account on Enterprise

Payload redaction

Mark fields as secret and they never leave your cloud

Uptime

99.95% SLA on Business, 99.99% on Enterprise

Get started

Ship the agent. Keep the receipts.

Free for 50,000 steps a month. No credit card, no separate workers, and your first durable workflow deployed before lunch.

Start building →

Book a demo →

$ npx orrindel init