Percolate An AI workflow engine in Postgres GitHub llms.txt

The AI-generated video, for which we apologise

Ten minutes on why any of this is in Postgres.

Percolate — the control plane. 9:51, captioned. mp4, 10 MB

What it covers

Eleven sections, in the order the pages here are written in. The argument is the one on what Percolate is, at the speed of somebody explaining it rather than the speed of a reference.

00:00 The experiment Push all of it down and look
00:52 What is left Everything here is a row
01:28 Built first Permissions and workflow, underneath the agent runtime rather than in front of it
02:23 Install a domain A domain, installed as a document
03:11 Sourcing Ingestion is a subsystem
03:59 Ask across modes One language over entity lookup, graph, semantic and lexical search — with plain SQL as the floor
05:33 Workflow semantics An agent turn is just a step kind
06:20 The slow participant When the next step is a person
07:08 Nobody holds the plan Postgres is the queue
07:29 What is in the box One SQL file, one small Rust parser, and the two processes that speak HTTP
09:34 Where the work is A database problem wearing an AI hat
Why it works — the video is the pitch, and the pages are the contract

Nothing here is evidence. The hosts assert things that the rest of this site either shows working or names as a cost, and where the two disagree the pages are right: they are checked against a running database on every build and the narration is not.

Read it as the argument for the design and then go and see whether the design holds — what it costs is the other side of it, and scaling is the side with numbers in it.