Transcription: OpenAgents Episode 032 - Toward Semi-Automation

Source: https://twitter.com/OpenAgentsInc/status/1731086330694651924 Wiki source: https://raw.githubusercontent.com/wiki/OpenAgentsInc/openagents/Video-Series.md Media title: OpenAgents - Episode 032: Toward Semi-Automation We review Faerie's pull request w... Upload date: 20231202 Transcription model: mlx-whisper/mlx-community/whisper-tiny Generated at: 2026-06-01T20:32:24Z

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

[00:00] Christopher David: Okay, we have a PR, let's review it.

[00:06] Christopher David: Update memory is controller and memory's test for credit functionality.

[00:10] Christopher David: Looks good.

[00:15] Christopher David: We don't actually have any changes to routes API and we kind of added a filter, I think.

[00:20] Christopher David: Okay, so here's the cool thing.

[00:22] Christopher David: We have test results in the PR.

[00:28] Christopher David: We have it being able to factor in my feedback and then writing the script again, it'll factor in those comments in and it'll make comments based on it.

[00:42] Christopher David: You can see that it's kind of like trying to do what I'm asking using a new memory factory.

[00:49] Christopher David: But the actual code has some of these before it like some of the convention that we're using for swapping in code gets kind of messed up.

[01:04] Christopher David: But the cool thing is that we have when we have tested fail, we get the result.

[01:14] Christopher David: And the test just can't run because there's a crash, we get that also in the PR.

[01:21] Christopher David: And if you look through some of my closed PRs here, there's a number of times when the test would pass.

[01:32] Christopher David: So we get that feedback.

[01:35] Christopher David: The only thing that's remaining for us to get this fully working is to tweak the prompts.

[01:45] Christopher David: Because it adding and leaving that like dash dash, just stuff, there's just sometimes when GPT 4 will give us the right answer it will do will we want other time to it will.

[01:57] Christopher David: So it was wrong or the filtering that we're using inner application to strip this stuff out doesn't match with the output is we need to add a little bit more deterministic responses and filtering.

[02:13] Christopher David: But we're at the point where I don't want to keep stabbing at this massive class that I've written.

[02:20] Christopher David: This is just sort of like horrible stuff, everything there and in the patcher file.

[02:28] Christopher David: We've sort of departed from test driven development, I've just been like pacing stuff in here and it's starting to smell.

[02:36] Christopher David: And one thing we're going to do both for you know developer sanity and building towards what we know that we're going to need in the product that will be live at openagents.com.

[02:50] Christopher David: Is inspectability.

[02:52] Christopher David: All of these steps that are being done by the agent where right now I'm just kind of like dumping all of the logging out on my console it's not very you know well structured.

[03:05] Christopher David: Hard to kind of inspect and debug exactly when things go wrong just kind of stabbing around but what we want to have is those kinds of things happening be the interface you can see each step of the agent process you can like see when it fails.

[03:23] Christopher David: Generally the idea of having an agent's thought process or every single thing that it chooses to do with the inputs now puts are at each stage will be important to.

[03:35] Christopher David: So we're going to kind of pivot from here to building out the UI a bit and our goal next is to have the kind of steps and decisions that are going on.

[03:49] Christopher David: You know visible here.

[03:52] Christopher David: The cool part about kind of splitting this up into the different like individual parts that will compose the final like agentic result is we're then able to do things like if you have a prompt that you want to suggest will increase the.

[04:09] Christopher David: Success rate of the agent because you figured out what the appropriate prompt could be we're going to allow people users to be able to.

[04:19] Christopher David: You know suggest improvements to various parts of the workflow or if we're using gbt for gbc.

[04:26] Christopher David: You know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you know, if you're going to do that, you

[04:56] Christopher David: know, so we're, you know, maybe I don't know, five or 10 videos away from kind of adding that more, you know, crowd sourced editability to this, but step one, let's start adding visibility. So we can get this to the point where when we run this script, it will give us something that's mergeable. We're obviously very, very close to that. Then we can experiment, hooking up to the scheduler and having it run and kind of full automatic mode.

[05:25] Christopher David: And letting us control that increasingly through our beautiful UI at openagents.com.

[05:32] Christopher David: Thanks for following along. See you soon.