marlow

About

There is no shortage of impressive agent demos and very little you would be willing to page someone about at 3am. Marlow is the second thing: every dispatch records which workflow ran, the agent that handled it, and what it was working from.

The job

Marlow reads tickets from Linear, Jira, GitHub or Slack — or a webhook you sign yourself. A ticket label naming one of the project's workflows decides what runs; a classifier steps in only when nothing is labelled.

The agent runs in its own sandbox, and the result goes back where the work already lives: a branch and a pull request, an artifact in your bucket, or a file on an SFTP host.

Principles

Agents are ordinary software
They need audit trails, retries and sandboxes for the same reasons every other production workload does. Nothing about a model changes that.
We do not profit from your tokens
You bring the model keys and your provider bills you directly — a margin on tokens would quietly reward us for choosing the wrong model.
Configuration beats prompting
A ticket label naming a workflow selects it outright, with no model call. Letting a model decide is the fallback when nothing is labelled, not the first move.
The console is the product
Most of the work is not the run itself, it is understanding a run that went wrong. If that takes more than a few clicks we have built the wrong thing.

Reach us

Email is the fastest way — every address is on the contact page.