Transcription: OpenAgents Episode 157 - Chains of Thought and Action

Source: https://x.com/OpenAgentsInc/status/1886297781138030777 Wiki source: https://raw.githubusercontent.com/wiki/OpenAgentsInc/openagents/Video-Series.md Media title: OpenAgents - Episode 157: Chains of Thought and Action We define a chain of thoug... Upload date: 20250203 Transcription model: gpt-4o-transcribe-diarize Generated at: 2026-06-01T06:51:38Z

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

[00:00] Christopher David: Welcome to the post-Deep Seek R1 era.

[00:03] Christopher David: Open source has caught up to the closed labs,

[00:06] Christopher David: and now we need to power past them.

[00:08] Christopher David: So we're introducing this concept called chains of thought and action,

[00:13] Christopher David: combining the magic of reasoning models like Deep Seek R1 with that same concept applied to agents,

[00:23] Christopher David: the tool use.

[00:24] Christopher David: Okay,

[00:24] Christopher David: this tweet got me thinking.

[00:29] Christopher David: blah blah blah my central problem with large reasoning models like o3 and r1 is that they claim to check their answer without really checking them why do all your thinking in a single generation if you can split it in 10 using much smaller modules yeah

[00:45] Christopher David: so this concept of reasoning generating a chain of thought like sure it improves over just like a single shot or few shot generation but but

[00:54] Christopher David: Who's to say that that thinking should be limited only to the training data?

[00:59] Christopher David: Why not give it access to live tools?

[01:04] Christopher David: Reality checks,

[01:05] Christopher David: yeah,

[01:06] Christopher David: like break this up into not just model calls,

[01:10] Christopher David: but also tool calls,

[01:11] Christopher David: you know, agents.

[01:15] Christopher David: So OpenAI had a release today.

[01:20] Christopher David: uh where does it go a little a little viral tweet here that's nice so the 175 157 billion soon 300 billion dollar company is posting model leaderboards where the top model deep

[01:38] Christopher David: research model this guy's from open ai by the way the

[01:43] Christopher David: top model

[01:45] Christopher David: Uses tools like browsing compared against models that do not use tools.

[01:52] Christopher David: This is beyond meaningless.

[01:54] Christopher David: It's deceptive.

[01:56] Christopher David: Look at the only thing on here that uses tools.

[01:59] Christopher David: Like that's not apples to apples comparison. This is shady.

[02:04] Christopher David: But what is deep research?

[02:06] Christopher David: It's a model,

[02:07] Christopher David: O3 or Bini or something with some fine tuning.

[02:11] Christopher David: that they give access to model to tools okay cool this should have been compared against some of the very many awesome projects that are out there many of them open source doing agentic web crawling um but then they would have only probably demonstrated an incremental improvement so um you

[02:33] Christopher David: know open ai not disclosing the reasoning traces

[02:38] Christopher David: they've said specifically in a previous blog post that they are intentionally hiding the chains of thought some people got excited at their the like reddit ama where altman yesterday was like oh they're like oh he's gonna show the chains of thought no they're gonna do better about maybe making improving it like just mealy mouth bullshit so we're in open source land here we're gonna show our traces we're not scared of whatever all this crap is

[03:05] Christopher David: Okay, but we want to see those for both reasoning as well as tools.

[03:15] Christopher David: So I figured that, you know, even though we may do our own little kind of archive post at some point, like the chain of thought folks did,

[03:23] Christopher David: you know, I'm just some dumb app builder,

[03:26] Christopher David: so I'm going to build a dumb little app.

[03:29] Christopher David: We're going to code this up.

[03:31] Christopher David: So here's the example.

[03:33] Christopher David: We're going to do a GitHub issue solver agent that submits a pull request with code that solves a given GitHub issue,

[03:40] Christopher David: combining multi-step reasoning and tool use to understand code bases and documentation,

[03:44] Christopher David: use CI tools and tests,

[03:45] Christopher David: and interact with the GitHub and other external APIs.

[03:50] Christopher David: I do a version of this every day. I do about 80% of my coding through the OpenAgents Pro dashboard. This is not really public-facing.

[03:58] Christopher David: just me and some kind of private clients but we're doing a version of this that's going to be public facing released later this month but just you know me saying stuff it loops through it shows me the tools that it uses input output and the next version of this is going to be a little bit more autonomous a little bit more like Devon but like full transparency for all these things okay so this agent is also and all of these agents should show the full trace of their Kota

[04:27] Christopher David: Including both the reasoning, using a model like DeepSeek R_ one, and tool use, displaying input-output of each tool as set to the LLM. Or if a model that's not a reasoning model, if it send it if sent for some reason to Claude, 'cause that's better at some things, still let's get some transparency about what are the inputs and outputs there. And let those be individually like tested or inspectable or upgradable.

[04:49] Christopher David: And so this workflow may combine various things like building a repo map for the issue context or previous video I was doing uh

[04:56] Christopher David: A repo map from ADER. We've since built our own version kind of modeled off of theirs internal to ROS.

[05:01] Christopher David: That's just a script.

[05:02] Christopher David: We're going to pass that script to some model or models.

[05:06] Christopher David: So one of the things I want to do in this next series of videos, I'll do a video on like building an eval so we can test to see what model or combination of models can generate the best and most relevant files based on the repo map relevant to the issue.

[05:23] Christopher David: We're going to traverse the relevant files,

[05:26] Christopher David: analyze things,

[05:27] Christopher David: plan out the changes,

[05:28] Christopher David: generate the code,

[05:29] Christopher David: generate a PR,

[05:30] Christopher David: and then do some sort of test in CI.

[05:32] Christopher David: So some combination of a bunch of different things and yeah,

[05:37] Christopher David: let's just be transparent and show all this.

[05:40] Christopher David: Um, blah, blah, blah. Yeah,

[05:41] Christopher David: if you're entrusting such an agent to make sensitive edits to a code base,

[05:44] Christopher David: if you're hiring an AI engineer and they don't show you their code,

[05:47] Christopher David: they don't show you their trace, their thought process,

[05:49] Christopher David: what they're doing,

[05:50] Christopher David: like, why would you trust that when you have an alternative called an open agent that does show you all that?

[05:57] Christopher David: All right, so we're going to build this issue solver agent over the next few videos of our series. Follow us on X if you haven't.

[06:06] Christopher David: It'll be fun.

[06:07] Christopher David: See you soon.