Blog
Notes from the people building the runtime.
Engineering deep dives, design notes and the occasional post-mortem. Written by the team, edited for people who have to ship.
All posts
Engineering
Retries are not durability
A retry repeats a call. Durability remembers what already happened. Why the difference decides whether your agent refunds a customer once or twice.
Engineering
How we keep p99 step overhead under 50 ms
Every step is a checkpoint. Here’s how we make checkpoints cheap enough that you stop thinking about them.
Product
Designing approvals people actually answer
An approval request is a tiny interface. Most are bad. Five rules we learned from 40,000 approvals.
Engineering
Fair queueing, explained with three customers
FIFO is fair to jobs and unfair to customers. A short, visual explanation of keyed concurrency.
AI
What an eval pipeline should cost
We cut our eval bill by 38% without changing a single prompt. The trick was making failures cheap.
Engineering
Sleeping for 14 days without a server
How step.sleep works under the hood, and why a sleeping run costs nothing.
Stories
The agent that refunded twice
A post-mortem we see every month, and the two lines of code that prevent it.
Company
Why we built Orrindel
We spent years building queues, retry tables and cron scripts at three companies. This is the runtime we wished we’d had.
Subscribe
One email when we publish.
About twice a month. Engineering posts only, and every issue has a reply-to that reaches the author.