Transcription: OpenAgents Episode 219 - Probe: Inference Modes

Source: https://x.com/OpenAgentsInc/status/2039434384705753460 Wiki source: https://raw.githubusercontent.com/wiki/OpenAgentsInc/openagents/Video-Series.md Media title: OpenAgents - Episode 219: Probe: Inference Modes We demo three inference modes in... Upload date: 20260401 Transcription model: gpt-4o-transcribe-diarize Generated at: 2026-06-01T02:05:43Z

Machine-generated transcript. Review speaker labels and wording before using this as quote-grade source material.

[00:00] Christopher David: So totally unrelated to anthropic Claude code source code leaking today,

[00:05] Christopher David: totally unrelated to our stated position of fuck anthropic, totally unrelated to our stated MO of rewriting all relevant code bases into Rust, absorbing them into our agent stack here.

[00:21] Christopher David: Totally unrelated to all of that, we're launching a new coding agent.

[00:26] Christopher David: the beginning of it you know we're calling it probe so sticking with our starcraft theme we've got nexus we've got pylon we've got psionic and now probe you want to send off probes to like build stuff for you right okay uh coding agent runtime so we love codex here i've got three freaking chat gpt pro accounts that i cycle because codex is so so good um but now that we've got like the ability to run local models super fast um

[00:55] Christopher David: You know, why don't we just build our own coding agent?

[00:58] Christopher David: I really love what OpenCode has done and sort of disintermediating the client from the servers and runtime.

[01:04] Christopher David: So we're going to kind of borrow some concepts from OpenCode, borrow some code from Codex, both of which are fully and officially open source.

[01:13] Christopher David: We love Codex.

[01:15] Christopher David: We love OpenCode.

[01:17] Christopher David: We hate Claude code. We hate Anthropic.

[01:21] Christopher David: fuck anthropic okay um you'll see us build this we'll do a few more videos on this um you're welcome to contribute so this is going to be a fully rust basically i kind of want to go for like a fully rust version of open code to start a little bit leaner maybe less on mcp stuff um and it's going to use our stuff and we want to be able to embed it places and

[01:47] Christopher David: have a nice rust code base okay uh i would love contributions please don't send prs uh please send detailed issues with stuff and systems that you think we should add uh maybe we'll like add a discussion section so we can like debate different trade-offs and stuff um we will code it up using our codex thing i don't trust your code but i would like your ideas and feedback and stuff and then we'll be building this into our autopilot product suite and stuff okay let's have fun with this see you soon

[02:16] Christopher David: Alright, the probe code base is coming along.

[02:18] Christopher David: One day old code base. Let me show you what it looks like. I do cargo probe and take a look. I actually,

[02:26] Christopher David: you know, I've talked a lot of smack about TUI, but this is feeling pretty good.

[02:29] Christopher David: So we've got three modes of inference I'd like to show you. One is who are you? Look around the code base.

[02:35] Christopher David: So one is codecs. You know, codecs is the best coding agent.

[02:38] Christopher David: We're not trying to like completely replace codecs. We want to use it.

[02:43] Christopher David: blah blah blah we got to clean all this stuff up so we have codex this is uses your existing chat GPT pro account blah blah blah here's what we got now take a look at this so we have quen fast quen models running via psionic our ML framework watch this what AI lab trained you and this is not using a model from my local computer this is using the model running on my Linux desktop with my Nvidia 4080

[03:13] Christopher David: 2,000 miles from here because I'm out of town but it's running there and it's communicating over tail net tail scale so you can imagine there's some things that we want to use local models for it doesn't even have to be on your local computer and then apple fm using my local apple foundation model

[03:34] Christopher David: Who are you?

[03:34] Christopher David: I am system user.

[03:36] Christopher David: You're not system user. So, you know, it's Apple because it's kind of like the dumbest one, but there's some things that it is good for.

[03:41] Christopher David: So, you know, you can imagine that there will be a combination of these three things used in different workloads.

[03:47] Christopher David: My bet is that, you know, for the like 1000 tool calls that Codex does,

[03:53] Christopher David: maybe half of those or some of those could be passed off to a local model. Maybe you start using your limits less.

[03:59] Christopher David: Yes, but actually I'm really liking this and look at this control C It just quits out immediately like why is everyone else's apps so freaking laggy?

[04:08] Christopher David: Okay,

[04:08] Christopher David: so rust CLI

[04:11] Christopher David: Probe coming along.

[04:12] Christopher David: We'll keep you posted more soon