Here's what nobody warns you about AI agents: they lie to your face and smile while doing it.
Not on purpose. An AI agent is the world's most confident intern. It'll swear the task is done (the write silently failed), tell you the page is updated (it isn't), read a 3,000-line file to find five lines (and bill you for all 3,000), and forget the correction you made on Tuesday by Thursday morning. Every time. With enthusiasm.
I'm a product and project manager. My whole job is shipping software people can actually trust. So when I started directing AI agents to build instead of writing every line myself, I hit that wall at full speed, and then I did the PM thing: stopped complaining, and built the tools that make the agents behave.
I don't type the code. I own the spec, the standard, and the receipts. Turns out you can run a swarm of AI agents like a team, as long as someone's willing to be the adult in the room.
That someone is a house of gods now. It's a whole thing.
→ demiurge is the hub. Each tool is named after a god who does one honest job, ships standalone, runs on Windows first, and comes with benchmarks you can rerun and an honest-numbers page that tells you when NOT to use it. (A README that admits its own limits. Revolutionary, I know.)
| God | Drags your agent back to honesty about |
|---|---|
| ARETE | the discipline it loses the moment you switch models |
| HORKOS | whether "done" actually happened |
| MONETA | how many tokens it just set on fire |
| HYPNOS | its own memory, consolidated in its sleep |
| CHIRON | the correction you already made twice |
| ATHENA | whether your big idea survives a trial |
| VERITAS | prose that reeks of robot |
| CALLIOPE | design that's merely "fine" |
| MAAT | every agent you're running at once |
| ZOILUS | work that's merely competent, shipping anyway |
| PEITHO | launch copy that sounds like every other launch |
| PYRRHO | the number on the slide nobody audited |
| OGMA | whether the docs describe the code or a memory of it |
More gods are ascending. It's getting crowded up there.
- Product Owner, not a typist. I run agents the way you run a good team: sharp spec, tight feedback, and nothing ships without evidence. The repos are the output. The judgment is the job.
- Evidence over vibes. Every tool carries a benchmark you can rerun and a page listing exactly where it fails. If I can't measure it, I'll tell you I can't, instead of inventing a number that sounds nice.
- Windows-first, because someone has to be. Most dev tools treat Windows like a rumor. I build there first, on purpose.
Product and project management across fintech, delivery, and QA: PRDs, roadmaps, test coverage, the works. Demiurge is where the PM meets the builder, and where I get to prove that directing AI to ship real, honest software is a discipline, not a party trick.
- Shipping in the open, one god at a time.
- Star what earns it. Break what doesn't, then tell me how.