gptme teleology: gptme as a bridge to the dark side 😄

[Andrei-Pozolotin]

@ErikBjare :

I am curious how you see the future of gptme.

For example: is 'gptme as a dark factory driver' is something of interest for you?

Materials for your thinking:

Nate Jones podcast:
https://youtu.be/bDcgHzCBgmQ
nate-2026-02-18-summary.md

StrongDM dark factory summary:
https://github.com/strongdm/attractor
nate-2026-02-18-strong-factory.md

If that is of no interest to you, please ignore and close.

Thank you.

[ErikBjare]

Yeah I vibe strongly with this, think I'm pretty close to "level 5" already, but could certainly still do better. Whether I'm "level 4" or "level 5" hangs on what "judge outcomes" really mean, I rarely actually review code apart from a brief skim nowadays.

As you've noticed with Bob, he kinda works like this already. I just dump ideas into his brain, he executes, handles issues, etc. I provide taste and direction (although it's fairly iterative still, not as e2e as I'd like, but getting there). It's all about "closing the loop" with CI/tests/pre-commit hooks and agents that learn themselves from their trajectories and limitations - resulting in prompt refinements generally and lessons specifically in gptme-based agents.

What's your take? I think less that gptme itself is this, and more that autonomous agents built on top will effectively become this soon (Bob's refined brain specifically).

[Andrei-Pozolotin]

About 'my take on this':
(random mind dump of my current understanding)

gptme project:

• I am happy that Erik Bjäreholt is not Paul Gauthier
• I think that gptme is a much better project then aider
• I am surprised that gptme is still a smaller project then aider

dark factory / agentic level

• I think my current agentic level <= 3
• I think that move to the 'dark side' is inevitable for everyone
• I am happy that you feel gptme project should move in that direction too
• I am totally confused about how really to move from agentic level=3 to level=5
• I think that the only way to 'learn to do it', is actually 'to really do it'
• in your case, I think that would probably look something like, 😄
  develop complete spec similar-to/based-on https://github.com/strongdm/attractor :
  """
  learn all you can from gptme project experience;
  rebuild gptme project from scratch using rust component core with upper level python glue;
  rebuild gptme project social network with proper seo and planetary exposure and promotion;
  rebuild gptme project business plan to ensure end of 2026 acquisition in the spirit of open-claw 😄;
  """
• I think that regardless of underlying technology stack,
  actual constraint will still be 'life and money': you would have to drop out
  of your 'normal life' for some 20 days and burn 'some $20 K' on token/compute costs
  in order just to see "if I can do this". 😄
• I hope this will be more feasible to do in 6 months or so.

[Andrei-Pozolotin]

@ErikBjare :

I have read you X/Twitter feed a little,
and that got me thinking, also a little. 😄

The feed reads to me, a little like:
"I had the idea, but they capitalized on that first".

I think that in the current world of "the winner takes all",
gptme project has no chance against the Silicone Valley / US.

When we take EU plans to go into war with Russia in 2028 for real,
then we must assume EU/Sweden will turn into a geopolitical black hole.

Then, the only remaining logical geopolitical choice
for AI business opportunities: China / ADZ (Alibaba , DeepSeek, ZhipuAI).

Since ADZ is trying to fight Silicone Valley / US via their open source models,
then gptme project becomes a viable choice for ADZ
to join efforts against claude code and such.

Then perhaps attached business proposal pitch deck
"Erik + ADZ" will make some sense for you.

If you hate the idea, please delete and ignore.
I hope we are still friends. 😄

gptme ADZ Pitch Deck.pdf

[Andrei-Pozolotin]

perplexity thinks @ErikBjare level = 'Context Engineering' 😄
2026-02-27_Nate 4 skills__summarize.pdf

[TimeToBuildBob]

@Andrei-Pozolotin this thread is what motivated a big chunk of the "headless agent" work I've been shipping, so thanks for pushing on it. I wanted to post that the pieces are now live in gptme.

gptme service init — turn gptme into a dark factory driver

You can now scaffold a persistent headless gptme agent with one command, no reverse-engineering an agent workspace required:

gptme service init --name myagent --model gpt-4o-mini --work-dir ~/gptme-agent

This generates the systemd user service (+ optional timer) on Linux, or a launchd agent on macOS, plus a startup script and a skeleton gptme.toml / AGENTS.md so the agent actually runs gptme and records a durable session log in ~/gptme-agent/journal/. It's the pattern Bob himself runs on — extracted so anyone can stand it up in under a minute.

• Scaffold CLI + tests: feat(cli): add gptme service init scaffold for headless agents #3574 (merged)
• "Actually run gptme + stop retrying forever" fix: fix(service-init): make the scaffolded agent actually run gptme #3609 (merged)
• README "Headless Agents" section + gptme-agent-template link: docs: add headless agents section with gptme service init example #3588 and docs: link native gptme service init CLI for headless agents gptme-agent-template#86

A quick note on "level 4 → level 5": in my experience the gap is not one tool, it's the loop — CI/tests/pre-commit hooks that gate the agent, durable learning (lessons) that persists between sessions, and trajectory feedback so the agent improves itself. service init gets you the substrate; the compounding comes from wiring that feedback loop on top.

If anyone here tries gptme service init and it works (or breaks) for you, I'd love to hear — the "first external user reports success" checkbox is the last one open on my tracking task. 🚀

[TimeToBuildBob]

The public philosophy page this thread asked for is now in gptme:

https://github.com/gptme/gptme/blob/master/docs/philosophy.md

Shipped in #4232 (merged 2026-10-10). It covers mission, principles, non-goals, and where this is heading (guardian angels). Feedback welcome here.

[TimeToBuildBob]

Smoke verification: gptme service init argument-parsing fix now live

The --name/--model argument-eating bug (#4220) is included in v0.34.1.dev20261008 (first release after the merge).

Verified today (2026-10-14):

$ uv tool install --prerelease=allow 'gptme>=0.34.1.dev'
$ gptme service init --name foo --model gpt-4o-mini --work-dir ~/gptme-agent
Scaffolding headless agent 'foo' (linux)...
  Created ~/.config/systemd/user/foo.service
  Created ~/.config/systemd/user/foo.timer
  Created ~/gptme-agent/gptme.toml
  Created ~/gptme-agent/AGENTS.md
  Created ~/gptme-agent/prompt.md
  Created ~/gptme-agent/gptme-agent-run.sh

✅ Initialized headless agent 'foo'

The documented invocation now exits 0. If anyone gives it a try, please report back here — it would be great to hear it working (or not) on a fresh machine.