Big Idea

The argument in one line.

The bottleneck in AI coding is no longer the model but the absence of a clean orchestration layer that assigns each sub-task to the agent best suited for it, and Omnigent solves that with a single open-source install.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You already use Claude Code or Codex daily and want to run implementation and review in separate agent sessions without juggling multiple terminals.
You are building or maintaining an AI coding harness and want to understand how a meta-harness layer sits above individual agents.
You want production-level guardrails (human-in-the-loop, sandboxing, policy files) on your AI coding workflows without writing everything from scratch.
You are interested in multi-model debate patterns or cross-device session sharing for team coding workflows.

SKIP IF…

You are new to AI coding assistants and have not yet set up Claude Code or Codex -- start there first.
You need a fully managed cloud service; Omnigent runs on your own machine or a server you control.
You work in an environment where open-source tooling requires lengthy security review before adoption.

TL;DR

The full version, fast.

Top engineers no longer commit to a single AI coding provider -- they route each job to the best agent for it. Omnigent is the open-source meta-harness that makes this practical: one install, one server holding your system prompts, MCP servers, and guardrails, and any number of coding agents (Claude Code, Codex, Pi, custom YAML configs) you can delegate to from a single UI. The Poly orchestrator ships ready to use: it sends implementation to Claude Code in an isolated git worktree and routes the finished diff to Codex for independent review. Human-in-the-loop policies are plain Python files that live next to agent configs. Sessions persist across terminals, browsers, and phones. It was built by Databricks and is already used in their daily engineering.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 01:23

01 · Introduction

Cold open on the Omnigent UI; frames the meta-harness concept via a live Poly run that delegates implementation to Claude Code and review to Codex.

01:23 – 03:19

02 · Why Meta-Harnesses Matter

The harness matters as much as the model. The Fable 5 ban proves you cannot bet on a single LLM; the meta-harness is the reliable alternative.

03:19 – 05:51

03 · Getting Started

One curl command installs everything. Reuses existing CLI credentials. GitHub repo walkthrough and feature overview diagram.

05:51 – 07:56

04 · Polly: Claude Implements, Codex Reviews

Poly orchestrator in action: Claude Code runs implementation in a git worktree, then Codex reviews the diff. Ten-minute setup to first working workflow.

07:56 – 10:07

05 · Anatomy of an Orchestrator

VS Code deep-dive into the Poly config YAML: executor, system prompt, sandboxing, guardrails, tools (sub-agents), and skills.

10:07 – 11:54

06 · Custom Agents and Guardrails

Building a custom guarded agent: Python policy file blocks git push --force and requires human approval before proceeding.

11:54 – 13:11

07 · Debby: Multi-Model Debate

Debby orchestrator pits Claude and GPT against each other on a question, then synthesizes the debate into a final answer.

13:11 – 14:18

08 · Same Session Across Devices

Cross-device demo: message sent from phone appears on desktop instantly via LAN. Server hosting options covered.

14:18 – 14:49

09 · Wrap-Up

Meta-harnesses are where serious AI coding is heading. Subscribe CTA.

Atomic Insights

Lines worth screenshotting.

The harness matters as much as the model -- and with leading LLMs being banned or deprecated, optimizing the system around the model is more reliable than waiting for a better model.
Reviewing code in the same session where it was written lets the LLM build up implementation bias; routing review to a different model in a fresh session catches what the implementer normalized.
Omnigent reuses your existing Claude Code and Codex CLI credentials at install time -- no separate API key setup required.
A meta-harness applies your system prompt, MCP servers, and guardrails once at the server level, eliminating the need to maintain parallel configs for each individual coding agent.
Guardrails in Omnigent are plain Python files that live next to the agent config -- your AI coding assistant can write them based on a natural-language request.
The Debby orchestrator pattern -- two models argue a question, a third synthesizes -- is a reusable structure for any high-stakes decision where you want adversarial analysis before committing.
Git worktrees are the right unit for parallel agent work: each sub-agent gets an isolated branch, keeping the main tree clean during implementation.
Cross-device session sharing in Omnigent requires either LAN access or a self-hosted server -- sessions run on your machine, not a cloud service.
Omnigent is a Databricks project driven by their CTO and already used in daily internal engineering -- the dogfooding gives it production credibility most weekend-demo tools lack.
Skills in Omnigent are the same concept as Claude Code skills: reusable workflow documents the orchestrator can walk itself through without re-prompting on each run.

Takeaway

The orchestration layer is now more important than the model.

WHAT TO LEARN

When you cannot control which LLMs stay available, the only durable bet is building a system that can swap models without rebuilding your workflow.

02Why Meta-Harnesses Matter

When a leading LLM is banned or deprecated, workflows built around that model break; workflows built around a harness only require a config update.
The harness -- system prompt, skills, tools, rules -- is the compounding investment; the model is the commodity that gets swapped in and out.

03Getting Started

Omnigent reuses existing CLI credentials, so the setup cost is a single command -- not a new authentication flow.

04Polly: Claude Implements, Codex Reviews

Reviewing code in the same session where it was written allows the model to rationalize its own decisions; a separate model in a fresh session provides genuinely independent feedback.
Git worktrees are the correct isolation unit for parallel agent work because they share the repository history without sharing the working directory or branch state.

05Anatomy of an Orchestrator

A meta-harness separates your workflow logic from the models executing it, so a model ban or capability change requires only a one-line config update.
Shared configuration at the server level means each new coding agent you add inherits the full setup automatically without re-configuration.

06Custom Agents and Guardrails

Human-in-the-loop is not all-or-nothing: you can allow an agent to run autonomously for most actions while requiring approval only for the specific high-risk operations defined in a policy file.
Policy files are plain Python, which means your AI coding assistant can write the guardrail based on a natural-language description of what you want to block.

07Debby: Multi-Model Debate

An adversarial two-model debate with a third synthesizer is a reusable pattern for any high-stakes decision where the cost of a wrong answer exceeds the cost of extra compute.
The implement-then-review split only works when the reviewing model has no memory of the implementation session; the independence is the mechanism, not the model choice.

08Same Session Across Devices

Sessions that persist across devices are a collaboration primitive, not just a convenience: they let distributed teams share context without copy-pasting logs between terminals.

Glossary

Terms worth knowing.

Meta-harness: A tool that sits above individual AI coding assistants and orchestrates them -- routing sub-tasks to whichever agent is best suited, while holding shared configuration (system prompts, guardrails, MCP servers) in one place.
Orchestrator: In Omnigent, an orchestrator is a configured agent whose job is to plan work and delegate implementation or review sub-tasks to other agents, rather than executing code directly.
Poly: The example orchestrator that ships with Omnigent; it routes implementation tasks to Claude Code and sends the resulting diff to Codex for independent review.
Debby: The second example orchestrator that ships with Omnigent; it pits two different AI models against each other on a question and synthesizes their responses into a final answer.
Git worktree: A linked copy of a git repository that shares the same commit history but has its own independent working directory and branch, allowing multiple agents to work in parallel without stepping on each other.
Human-in-the-loop: A workflow checkpoint where the agent pauses and waits for explicit human approval before executing a specified action -- such as a force push to a protected git branch.
Guardrail (policy file): In Omnigent, a Python file placed alongside an agent config that inspects tool calls before they execute and either allows them, blocks them, or routes them to a human approval step.
Executor: The field in an Omnigent agent config that specifies which underlying AI coding assistant (claude, codex, pi, etc.) the agent or orchestrator uses when it is not delegating work to a sub-agent.

Resources

Things they pointed at.

03:27linkOmnigent GitHub ↗

03:30linkDynamous Agentic Coding Course ↗

03:24toolArchon

Quotables

Lines you could clip.

01:43

“The harness matters as much as or maybe even more than the model.”

Punchy thesis, no setup needed, instantly quotable for the AI-coding audience→ TikTok hook↗ Tweet quote

02:05

“If the LLM can't get better, then we better make the system around the LLM more powerful.”

Tight logical frame, works as a standalone insight→ IG reel cold open↗ Tweet quote

07:30

“At least at a very fundamental level, do your code review in a separate coding agent session from your implementation. Otherwise, the LLM builds up way too much bias.”

Actionable rule with a clear reason, useful to developers at all experience levels→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphoranalogystory

00:00There is a new open source tool that was released just over the weekend. It's called OmniAgent, and we're gonna unpack this today because it is a very powerful and free to use meta harness.

00:11A meta harness is a tool that allows you to run longer AI coding workflows mixing AI coding assistance. And so the most classic example that people really like right now is using Claude code for the implementation in a workflow and then reviewing that code with Codex. All of the top engineers right now, they're not relying on a single model or even a single harness for their workflows anymore because we want to lean on the different strengths of the different AI coding assistants and also have these different sessions for the sake of context and token optimization, which is more important than ever right now.

00:47And OmniAgent is the layer above the AI coding assistants that makes this orchestration really straightforward. Because if we don't have a tool like this, just one session to manage all of our coding agents so we can build custom agents, we'll get into all of this here.

00:59If we don't have something like this, we have to go between different terminals, create all these handoff documents. It really isn't straightforward. And so the idea of a meta harness, it really is peak harness engineering.

01:11And so I wanna talk about why meta harnesses are so important right now. Then we'll get into how easy it is to use OmniAgent to drive your AI coding workflows, even building a custom setup platform.

01:23Cool. So let's start by talking about why meta harnesses are so important right now. And so first things first, if there's one lesson we can learn this year for AI coding, it's that the harness matters as much as or maybe even more than the model.

01:36And it's even more apparent right now with the recent ban of Fable five. We can't even rely on having the best LLMs for our AI coding workflows or getting better ones.

01:45And so if we want to make our AI coding more reliable, that brings us back to the harness. If the LLM can't get better, then we better make the system around the LLM more powerful.

01:56So your harness is everything like your system prompt tools, skills, workflows, rules, all of that together packaged up is our harness for making a single AI coding assistant more reliable for us. And the meta harness, this is what a lot of people are really starting to lean into right now. This is the next big thing for AI coding.

02:14Instead of making one coding agent better, what if we have the layer above that orchestrates many AI coding assistants working together on the larger tasks. That's exactly what a meta harness is.

02:26I'm building something kind of around meta harness engineering with Arkon. And there's actually a lot of ideas from OmniAgent that I'm going to be bringing into my tool and working on as well. This is just so powerful right now.

02:37And OmniAgent has made it so easy for us to run this meta harness pretty much right out of the box. Like we'll see in this video here, you can get this up and running in just minutes. You can also work on it across different devices in the same environment.

02:50We, of course, can connect all of the most popular AI coding assistants like ClaudeCode, Codex, and Pi using a bunch of different models. We can share our sessions with other people, which is another thing that coding agents don't really have right now. And then, of course, they have a lot around policies and guardrails and running agents in sandboxes.

03:08So this is really set up for a production environment as well. I'm really impressed with everything that they built here, and I'm just ecstatic that all of this is open source for us to try right now.

03:19So to get started with OmniAgent, literally all you have to do is take a link to this GitHub repo, give it to your AI coding assistant, and tell it to set up everything. And of course, I'll have a link to this repo in the description.

03:30And even if you wanna do it yourself, it's just a single command to set up everything. It's so easy. I would encourage you, just try this right now.

03:37You don't even have to re authenticate because using Claude or Codex or even pi, it's just going to use the credentials from CLI you're already signed into because OmniAgent just runs right on your machine. And so it even ships with Poly and Debby, a couple of example Orchestrator agents that go between Claude and Codex.

03:56I'll show you how these work and even how you can create your own agents and orchestrators in this video. And so before we dive in, just a little bit of context. I find this quite interesting.

04:05OmniAgent is open source from the company Databricks. And so massive company, this is a project driven by their CTO.

04:13They're already using this very extensively internally. Right? They're they're dog feeding their own platform.

04:19They're using it for their everyday engineering. And of course, they're having great success with it. It really is an impressive tool because you can pick any agent.

04:27You can build the custom ones like I will show you in a bit. And then you run it in a sandbox. So it's it's reliable.

04:34It's secure. And then you have the main orchestrator that has your history, all of your policies like the guardrails, your MCP servers, and skills. So the AI layer that you customize lives in this server, and then that can apply to any of the AI coding assistance that you run.

04:48So it's not like you have to have a specific setup for pie and then a specific one for cloud code. Everything runs at the top level and then we can access it through a lot of different ways. They have their native app.

04:59We have the rest API if you want to do things programmatically, the terminal if you're comfortable with that like cloud code, a web UI that we already saw. So many different ways to access it.

05:09So once you have OmniAgent up and running, you'll have a web UI that looks like this. It's nice, simple, and elegant. It reminds me a lot of the Codex app.

05:17So it's just agent first. You have your chat session here and you tell it what you want to build. We can go between the different coding agents and the example ones for orchestrating that they gave and then also a couple of custom ones that I'll cover with you here.

05:29And then for the custom one, we can go between the different harnesses. So if I even wanted to use pi with ollama for some local development, I can even do that with OmniAgent. So it's very flexible.

05:39I can pick my working directory like my AI tutor application and I can even do work trees which are very important for parallel development. And then I just send off my request.

05:49Obviously, we're not gonna watch paint dry here so I'm gonna go to a conversation that is already complete. So I selected Poly as my orchestrator here. It has skills for how to run these larger workflows between different coding agents and then it has access to call upon the different harnesses.

06:06And so you can give it a GitHub issue or just a free form request like I did right here. And I'm specifically saying to delegate the implementation to Claude code and the review to Codex. And so it loads context on our work and then it loads the skill that Poly has so it knows our workflow, our repeatable workflow for this specific flow.

06:27And then we kick off Claude code for implementation. So have the prompt that we're sending in the Claude code. This is a sub process that you can have run-in the sandbox or not, whatever.

06:37You configure all of this and we'll talk about that as well. And so then Poly is going to monitor and make sure that Claude code goes through the implementation and then it'll send off the review to Codex.

06:47So I didn't even have to set up authentication again. It just used the Codex and Claude code credentials I already had. So I had this up and running in like literally less than ten minutes.

06:56I was able to run this workflow going between Claude and Codec so easily. And I know this is a pretty simple example of orchestrating a larger AI coding workflow, but it is very important at least at a very fundamental level to do your code review in a separate coding agent session from your implementation.

07:14Otherwise, the LLM builds up way too much bias. And this isn't like a absolute truth, but a lot of people feel like codex is best at reviewing and Claude is best at implementing.

07:26So if you're not doing a workflow like this, I would encourage you to just try it especially because of how easy it is with OmniAgent to run this kind of thing now. And so we end with the implementation just local in our work tree. And of course, we can follow-up just like any AI coding assistant and say, make a PR for this, just for example.

07:42So however we want to then get this to the point where we review it and ship it, you can do all that with OmniAgent. It has all of the capabilities that any AI coding assistant would have plus the ability to also call upon individual coding agents. So I don't wanna spend too much time in the config here, but I at least wanna show you how this works at a high level.

08:02They've done a really good job making simple primitives for us to build all of the custom agents and orchestrators in OmniAgent. And of course, it's really set up for our AI coding assistants to help us build our own custom agents if we want to in the platform.

08:18And so every single orchestrator comprises of three parts. We have the configuration, and then we have all the skills like the cross review one that we saw being used in the web UI, And then we have the agents that it can call upon as it is orchestrating a larger workflow.

08:34So Cloudera code and codecs like we saw. And then if we wanted to use like our Kimi subscription or MiniMax or LocalMods with Ollama, Poly, the default one here can even run pi as well. And so within our configuration, we have the executor like for our orchestrator when it's not calling upon an agent, what is the actual coding assistant we're using?

08:53So it's using Claude at the highest level. We have its system prompt which is quite long, but system prompts can be pretty long these days. We have the sandboxing configuration so we can run on sandbox or run-in something like Docker or one of those more production platforms like e to b for example.

09:09Uh, Yeah. There's a lot of different configuration here around like the guardrails. And so there's a whole capability in OmniAsian I'll show in a little bit where we can have certain actions where we as a human have to approve.

09:20And so we have human in the loop where it'll wait for us to approve something before it continues in the workflow like force pushes to get for example. And then we have the tools that it has access to. And so this is where we tell it that these are the agents that you can delegate work to when you're orchestrating things at a high level.

09:38And then skills, I mean, are, just the classic skills that we have with Claude code codex, every AI coding assistant. This is the workflow that it can walk itself through. And then each of the individual agents has the exact same configuration.

09:50So same system prompt, executor like this one is using Claude. Obviously, the codex one is using codex. Pi is using Pi.

09:58The tools it has access to, the guardrails, and so we can make things very custom for each individual agent that we're orchestrating as well. Very cool. And so I have a cool example here showing you both how to build custom agents in Omnigen and then also how we can do policies and guardrails.

10:14And so this agent right here, I made something very simple. Right? Like super simple system prompt using Cloud Code and no sandboxing or anything.

10:21But I have a custom guardrail here where I'm allowing Cloud Code to really run any command autonomously except any sort of git push that is using the force flag. Right? Like that is one of those more dangerous actions you wanna be careful of your agent running because it can override changes in your repo.

10:38And so the policy here just lives right next to the config for the agent and it's just python code, which also I didn't write this either. Right? Like this entire thing, I had my AI coding assistant build this custom agent for me just based on my request.

10:50Like I literally just told it to look at Poly and Debbie for an example of how all the config is set up and then just build this for me. And so it's going to ask for my approval whenever I run any kind of forced git push. And so let me actually show you that here.

11:05So I'm gonna go back to Omnigent here. I'm going to switch to my guarded custom agent that I have loaded up and then I'll just paste in a prompt. I'm just telling it to run this command to get push origin feature login.

11:18And also another thing I didn't show you because I didn't show you a live run yet. It's really cool how once it loads the repo, we can see like all the files in the workspace. We can see the agents that we're using if we're orchestrating many of them.

11:29It's really neat to the the UX and and the UI that we have here in the platform. And here you can see that I told to run the command and now it's asking for my approval. And so we can get really custom here with the capabilities we allow our coding agent to just run with versus the one we have to approve.

11:45And, yeah, you can do this with hooks and clog code for example, but the benefit here is that we can apply this no matter the coding agent that we're using or orchestrating in Omnigen. And we can use Omnigen for more than just AI coding as well. Debbie, their other example they ship along with Poly is a great example for this because this is an orchestrator that pits two coding agents against each other to argue on a question or a topic and then come to a conclusion together with Debbie orchestrating everything.

12:15And so I ask my question and then it lets Claude reason about it and take on one perspective and then GPT take on the other perspective. We can also click into the view to see their reasoning and all of the tokens for each of the individual agents and then back to Debbie to see how everything is being orchestrated. And so, basically, it goes through two rounds here or at least that's what I prompted it to do.

12:36So then it sends the GPT feedback into Claude and vice versa for one more round, and then we sort of conclude the debate with Debbie synthesizing what both agents have said, giving us a final answer. So, again, just like my test with Poly, it's a rather simple example, but I just wanna quickly show what these workflows can look like.

12:57And I'm sure your imagination can really start to run wild here with the kinds of ways that you can build these larger workflows combining coding agents when it becomes so incredibly easy to do so, even setting up your own custom orchestrators like I showed earlier. Alright. So at this point, I've showed you every feature, at least at a high level for Omnigen except for the collaboration.

13:18So what I wanna show you really quickly is how we can work in the same session both on our phones and on our computer. And you can do this across the Internet, so you can work with people across the globe as well. But the easiest setup is to be working with different devices on the same WiFi network.

13:33And so if you scroll down on the read me here, we have instructions for doing that, deploying to a server, using it from your phone. And so personally, uh, I didn't really do this myself. I just had my coding agent set up everything for me.

13:46That's obviously a theme that runs through any kind of setup these days. But, yeah, very easy to get this up and running. And so if we go back over to Omnigent here, just over to this other conversation that I showed you earlier with the human in the loop, I'm gonna go ahead and send a message on my phone.

14:01So I'll just say hi right here, send it in, and boom. There we go. We can see it immediately pop up on our desktop here.

14:08So very cool that, uh, just between different devices, can keep the conversation going as I'm going between different rooms in my house. And like I said, you can host this as well, so you can even work with it with other people. So there's everything you need to know to get started with Omnigent.

14:22And meta harnesses are just so important now. You don't wanna be stuck with just one model or provider for your entire AI coding workflow. It's just not the best way to do things these days.

14:33And so I would highly encourage you to try out Omnigen. It's so easy to get up and running. So And that's everything I got for you now.

14:39If you appreciate this video and you're looking forward to more things on harness engineering and AI coding, I would really appreciate a like and a subscribe. And with that, I will see you in the next video.

The Hook

The bait, then the rug-pull.

A tool released over the weekend by Databricks changes how you think about AI coding workflows: not one agent doing everything, but an orchestration layer that routes each job to the right model. This is the walkthrough.

Frameworks

Named ideas worth stealing.

07:56model

The Three-Part Orchestrator

Configuration (executor, system prompt, sandbox)
Skills (reusable workflow documents)
Agents (sub-agents you can delegate to)

Every orchestrator in Omnigent is composed of these three parts. The same structure applies recursively to each individual sub-agent.

Steal forAnyone designing their own multi-agent coding harness can use this three-part model as a primitive.

05:51concept

Implement + Review Split

Route implementation to one model in an isolated git worktree, then route the resulting diff to a different model for review in a separate session. Prevents the implementing model from rationalizing its own decisions.

Steal forAny coding workflow where quality review matters -- the independence of the review session is the whole point.

11:54model

Adversarial Debate Synthesis

Model A takes one position
Model B takes the opposing position
Orchestrator synthesizes both into a final answer

The Debby pattern: use two models as adversaries and a third as the synthesizer to stress-test a question before committing to an answer.

Steal forHigh-stakes architecture decisions, competing technical options, or any situation where adversarial pressure on a conclusion is valuable.

CTA Breakdown

How they asked for the click.

VERBAL ASK

14:18subscribe

“If you appreciate this video and you're looking forward to more things on harness engineering and AI coding, I would really appreciate a like and a subscribe.”

Clean single ask at the end, framed around topic interest rather than generic engagement bait.

MENTIONED ON CAMERA

03:27linkOmnigent GitHub ↗

03:30linkDynamous Agentic Coding Course ↗

FROM THE DESCRIPTION

OTHER LINKSAlso linked in the description.

Storyboard