Hermes Agent + Codex: How to Use Codex as a Runtime Monitor
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I go through a real Hermes Agent user story. A builder used Codex to watch Hermes while it ran agent-to-agent work, catch the spots where it broke, and fix them live. This one is for anyone running AI agents (Hermes, OpenClaw, or something else) who is tired of workflows that fall apart halfway and wants a pattern they can copy.
I recorded this in May 2026. These tools change fast, so check the current docs for Hermes Agent, Codex and OpenClaw before you set anything up.
What I cover in this video
- 0:00: Why a Hermes + Codex workflow matters for OpenClaw users, and why reliability is the missing piece
- 1:32: The source, a user story on the official Hermes Agent user stories page from Nous Research
- 4:33: How agent-to-agent workflows actually break, step by step
- 5:18: Codex as the runtime monitor, and why that works the way a real engineering team does
- 6:48: Why Hermes is a threat to OpenClaw, though not for the reason most people give
- 9:04: The watcher-agent pattern, plus how I use Claude Code to fix my own OpenClaw setup
- 11:21: How to find where your workflow breaks and where to put the monitor
- 13:38: What to do next on Hermes, on OpenClaw, or if you're brand new
The Hermes + Codex story
The source is the official Hermes Agent user stories page from Nous Research. One builder, @gkisokay on X, was on day 10 of a build for their Hermes agent and wrote that Codex "saved the day as a runtime monitor" for their agent-to-agent workflows. They ran Codex with GPT-5.5 on extra high. Codex watched the workflow run, found where it broke, and fixed it live until it worked reliably.
That might sound like a small detail. It isn't a bedtime story or an inbox summary. The builder is using a coding agent as a reliability layer on top of another agent system. I think that's where serious agent building is going.
For context, the description notes that when the GitHub API was checked on May 20, 2026, the NousResearch/hermes-agent repo had 158,870 stars, 25,755 forks and 12,409 open issues, and had been pushed that same day. Stars aren't users or revenue, and they don't mean a tool is stable. When a repo is moving that fast and builders are posting real workflows from it, though, it's worth studying.
Why agent workflows break, and what a monitor does
The real problem with AI agents isn't intelligence anymore. It's follow-through. Can the agent keep its state in check, notice when a workflow broke, recover without you sitting there watching, explain why it failed, and improve the same workflow instead of starting from scratch each time?
Here's how a typical agent-to-agent workflow goes wrong. One agent plans, another executes, another checks something, and another uses a tool. The first tool call works, then the second. The third returns something weird, and the agent misses it. The next agent inherits that bad state, the whole workflow drifts, and you end up with a confident answer built on a broken middle step.
That's the job Codex was doing here. It watches the run, sees where it breaks, helps patch the issue, and the workflow runs again until it's reliable. It looks a lot like a real engineering team, where one person builds, one reviews, one watches the logs and one catches the broken assumption. Now some of those roles can be agents.
Business owners don't care which tool has more GitHub stars. They care whether the customer email goes out, the report is right, the lead list is clean, and the agent can fail safely without embarrassing them. That's why I think the agent market will be won by reliability: logs, checks, approvals, recovery, memory, permissions and repeatable workflows.
The watcher-agent pattern OpenClaw users should copy
Hermes is a threat to OpenClaw because it's pulling builders into serious experiments with multi-agent systems, local machines, coding agents and memory. That's OpenClaw's home turf. Attention turns into tutorials, then community, then plugins, then default behavior. (I go deeper on this in OpenClaw vs Hermes.)
The lesson isn't "use Hermes instead of OpenClaw," though. It's this: stop trusting one agent to do everything alone. You need a second agent that can read terminal output, inspect files, watch logs and tell you where things are falling apart. That could be Codex watching Hermes, Codex watching OpenClaw, or OpenClaw orchestrating Codex runs.
In my own OpenClaw setup, I always have a Claude Code terminal pointed at my OpenClaw workspace folder. When something breaks, I tell it something like:
Check out my OpenClaw setup. Something's not working. The gateway's not responding. The agent's lost track. My Telegram's not connected.
It goes through the files and fixes the problem. That's the same idea. To be clear, I haven't tested the exact workflow from that post, but I run something very similar. And I'm not saying Hermes is suddenly more reliable than OpenClaw.
Here's the pattern:
- The builder gives a goal.
- An agent does the work.
- A monitor watches the run.
- The monitor catches the drift.
- A human approves the fix.
- The system records the lesson.
- The next run gets better.
Where to put the monitor and how to compare tools
Ask yourself where your current workflow breaks. Does it break when the browser returns something unexpected, when a tool times out, when the agent writes code but never runs the tests, or when it forgets the original goal after 10 steps? That's where the monitor goes.
Start with one boring workflow that actually matters, like drafting inbox replies, turning customer calls into action plans, scanning failed builds or researching content topics. Then add a second agent whose only job is to check the first one. Did it finish the task? Did it use the right source? Did it make a claim it can't prove? Did it write files but skip the rest? Does anything need human approval? The watcher isn't there to be creative. Its job is to say "this looks broken."
This also changes how I'd compare Hermes and OpenClaw. Don't ask which one is smarter. Give both the same job, the same repo, the same approval rules, the same memory requirement and even the same failure case. Then see which one recovers better, explains failures better, gives cleaner logs, respects permissions, makes it easier to add a monitor, and turns a one-time fix into a reusable skill.
For next steps: on Hermes, give it one repeatable workflow and pair it with Codex as the watcher, inspecting the logs, output, state and final result. On OpenClaw, use OpenClaw as the operator and Codex as the code workflow reviewer, or have OpenClaw trigger a Codex subtask and verify the output before anything goes public. If you're new, don't start with 10 agents. Start with one worker and one watcher.
Key takeaways
- A Hermes Agent user used Codex as a runtime monitor to catch and fix breaks in agent-to-agent workflows until they ran reliably.
- The hard problem with agents now is follow-through and reliability, not intelligence.
- The pattern matters more than the brand. It can be Codex watching Hermes, Codex watching OpenClaw, or Claude Code fixing your OpenClaw workspace.
- Compare agent tools on the same workflow and the same failure case, not on which one seems smarter.
- Start small with one worker and one watcher on one boring workflow that matters, and think like an operator rather than a prompt engineer.
FAQ
How do you use Codex as a runtime monitor for Hermes Agent?
In the user story I cover, Codex watched the Hermes agent-to-agent workflow while it ran, caught where it broke, and fixed it live until it worked reliably. The builder ran Codex with GPT-5.5 on extra high. My advice is to give Hermes one repeatable workflow and have Codex inspect the logs, output, state and final result.
Is Hermes Agent a threat to OpenClaw?
I think it is, mostly in terms of attention. Hermes is pulling builders into serious multi-agent experiments, which is OpenClaw's home turf. But I'm not saying Hermes is more reliable than OpenClaw, and OpenClaw users can copy the same watcher pattern.
Can I use the watcher-agent pattern with OpenClaw?
Yes. You can use OpenClaw as the operator and Codex as the code workflow reviewer, or have OpenClaw trigger a Codex subtask and verify the output before anything goes public. I also keep a Claude Code terminal pointed at my OpenClaw workspace folder to diagnose and fix breaks.
How many agents should a beginner start with?
Start with two: one worker and one watcher, on one boring workflow that actually matters. Once you understand the pattern, build up from there.
Chapters
Transcript
Read the full transcript
0:00In this video, I'm going to be talking about the Hermes agent workflow that should make Open Claw users pay attention. Not because Hermes magically replaced Open Claw overnight, and not because one screenshot means the whole market has shifted to Hermes, but because a real builder used Codex to watch Hermes while Hermes was doing agent-to-agent work. So, integrating Codex with Hermes is the unlock here. It catches the breaks, and it fixes the workflow until it became reliable, which is kind of the Open Claw flaw at this point. Things break, it gets annoying, and then you give up. With Hermes and Codex paired together, this solves a lot of issues. So, that is the missing part to all these agentic workflows is
0:46reliability. Most people are still thinking about AI agents like they're chatbots, and we're way past that. You know, you type in a task, the agent does the task, and maybe it succeeds, or maybe it gets lost half the time. Maybe it even burns through tokens and gives you a half-baked result that looks confident, but doesn't actually work and ends up costing you thousands of dollars. That is not where this is going. The next phase is AI agent supervision. One agent doing the work, another agent watching that agent do the work, and then another agent processing and checking the logs. The human stops babysitting every single tiny step and starts designing the system. That's the shift that I'm trying to talk about in this video. Hermes is starting to show up in that lane in a real way. The
1:32source for this is the official Hermes agent user stories page from New Research. I have it linked down below, you need to check it out. They have a story from a builder named Gus Alke on X, and the story says, "Of day 10 of building an AGI for my Hermes agent, Codex saved the day as a runtime monitor for my agent-to-agent workflows. He used Codex with GPT-5.5 on extra high to watch the workflow run, and it caught it where it had broken and fixed it live until it worked reliably. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our
2:17community down below, Shipping School. We have a full Claude code course, a full open Claude course, and four live boot camps every single week where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you could book a call with me, we could share screens, and I can help you get Claude code or open Claude running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago, and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below.
3:02Get in now before the price goes up. That sounds pretty small until you understand what it means, and if you've ever had AI agent workflows break halfway, you know exactly what I'm talking about. It's very frustrating. This is not Hermes writing a bedtime story, either, and this is not Hermes summarizing an inbox. This is not another example of someone asking an agent to make a landing page. Way further than that. This is a builder using a coding agent as a reliability layer over another agent AI system. And that is exactly where AI agents and builders are going. Because the real problem with AI agents is not intelligence. Not anymore it isn't. The real problem is follow-through. Can it keep the state in check? Can it notice
3:47when a workflow broke? Can it recover without you staring standing there at the computer waiting for it to finish. Can it explain why it failed? Can it keep improving the same workflow instead of starting over every time like it has amnesia? I know you've been there before. It's frustrating. That is the whole game, and that is also why Hermes is becoming a real viral lane for Open Claw. They're competitors, but they make each other better, which is amazing. When I checked the live GitHub API today, new research / Hermes agent was sitting at 160,000 stars, 25,000 forks, and over 12,000 open issues. It had been pushed the same morning. Now, stars are not users. They're kind of like a almost
4:33like a fake kind of vanity metric at this point, and stars are not revenue, and stars do not mean the tool is stable at all. But, when a repo is moving that fast, it has much of the community's attention, and real builders are posting workflows like this, and you can't dismiss the hype. You have to study it. Here is the actual workflow pattern. Hermes is running some kind of agent-to-agent workflow, right? Maybe you have one running right now. That means there are multiple moving pieces, whether they're scheduled on a cron job, or they happen in the chat. Maybe one agent is planning, maybe one agent is executing, maybe another agent is checking something, maybe another agent is using some type of tool. So, if you have ever built anything like this, you know exactly where it breaks. The plan sounds good, and the first tool call
5:18might work, the second tool might work, maybe the third tool call returned something weird, and you start to notice the agent misses it. Then the next agent inherits the bad state. Then the whole workflow starts drifting. By the end, you get a confident answer built on a broken middle step. That is a nightmare. So, this builder used Codex as the runtime monitor. Codex watches the run. It sees where the workflow breaks, and then it helps patch the issue. Then the workflow runs again until it becomes reliable. This is not a chatbot trick, far from it. This is a closer to how real engineering teams work. This is a one-person build, one person reviews, one person watches the logs, and then one person catches the broken assumption. Except now, some of these rules can be agents. They don't need to
6:03be people. And that is where the money is. Because normal business owners do not care whether Hermes has more stars than Open Claw. I mean, come on, what are we talking about here? They do not care what model is under the hood. They care if the workflow actually runs reliably. They care if the customer email gets sent. They care if the report is right. They care if the lead list is clean. And most importantly, they care if the agent can fail safely without embarrassing them. This is why I keep saying the agent market is not going to be won by the flashiest demo. Come on now. It is going to be won by reliability. And reliability usually looks pretty boring. Logs, checks, approvals, recovery, memory. Who cares? Even permissions and repeatable workflows. Who cares? The stuff nobody
6:48wants to talk about until their agent makes a mistake in front of a customer. I care about that. So, Hermes is a threat to Open Claw. I agree. Yes, it is. But it is not in the simple way people frame it. Hermes is not a threat because one person posted some cool workflow. Hermes is a threat because it is pulling builders into serious experiments. It is making people try multi-agent systems. It is making people connect local machines, coding agents, workflows, and memory. And it is making people think about agents as an operating system for your work. This is Open Claw's home turf. So, if Hermes gets the attention and Open Claw only gets reliability nerds, that matters. Attention becomes tutorials, tutorials become community, and community becomes plugins. Plugins become default behavior at this point, and default behavior
7:34always wins. But here is the crazy take of all this. The workflow is not only a threat, it is also a pattern, open claw users should copy immediately. If you're struggling to keep up with content, well, I'm about to save you about 40 days worth of work. [music] I built something called content machine. It's 10 AI agents that run on the open claw orchestration, and they handle everything. Scripts, thumbnails, X posts, blogs, outreach, clips, newsletters, [music] all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning I wake up, and the content's already done. I spend maybe 15, 20 minutes reviewing and
8:19approving them, and I move on with my day. It works for any niche, fitness, finance, real estate, marketing, whatever you are building, and it is 100% completely customizable to your use case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the community. So, you plug in your own thing, and it molds it to you, it learns how you talk, and [music] it writes so it doesn't sound like AI slop. $97 one time, it's not a subscription. I'll put the link down below, and you'll thank me later. Because the lesson is not the use of Hermes instead of open claw, the lesson is stop trusting one agent to do everything alone. If you are building a
9:04serious workflow, you need a watcher agent. You need a second brain watching the first brain. You need an orchestrator agent that reports back to an a regular agent, and they report back to the orchestrator agent to check the work. You need something that can read the terminal output, inspect the files, watch the logs, and tell you where the workflow is falling apart. That could be Codex watching Hermes. That could be Codex watching open claw. That could be open claw orchestrating Codex runs. For example, in my open claw setup, I always have a Claude Code terminal pointed at my Open Claw workspace folder. If anything breaks in Open Claw, I just go to my Claude Code terminal and I say, "Check out my Open Claw setup. Something's not working. The gateway's not responding. The agent's lost track.
9:51My Telegram's not connected." And it goes through all the files and fixes it. This is the same concept. You know, that could be a separate quality agent, right? That only verifies outputs, but I'm just trying to explain what how I have my setup going. No different here between Codex and Open Claw, working with Open Claw and Codex Open with working with Hermes. And really, the brand matters less than the pattern here. It's just about the concept. The pattern is very simple. Builder gives goal, agent does the work, monitor watches the run, and monitor catches the drift. Human approves of the fix, system records the lesson, and the next run gets better. That's why Hermes is awesome. That self-recursive feedback loop to make things better, to get better at the skill and the task. And that is the direction I'm talking about
10:36here. Want to be very clear about something, too. I'm not personally tested this exact Giskard workflow from the post, but I run something very similar. And I'm not claiming every Hermes user can set this up in 5 minutes, and I'm not saying Hermes is suddenly more reliable than Open Claw. What I'm saying is the source of this is real, and the pattern is real, and the market signal is real. Builders are moving from single agent demos to supervised agent systems. That's the important part here. And if you're inside the community, inside Shipping School, or just building with Open Claw in general, this is the kind of thing that you should be studying. Not the logo, not the drama, the actual workflow. Ask yourself this: Where does my current agent workflow break? Does it break when the browser return something unexpected? Does it break when the tool
11:21times out? Does it break when the agent writes code, but never runs the test? Does it break when it forgets the original goal after 10 steps? Does it break when the model tries to be helpful instead of being act That is where you put the monitor agent and it does not have to be complicated. Start with one workflow, a boring workflow at that. Something that actually matters. Maybe it checks your inbox and drafts replies. Maybe it turns into customer calls into actual action plans like we do inside of our community. Maybe it scans failed builds. Maybe it researches content topics on X or whatever. Maybe it uploads to CRM or whatever. Then add the second agent whose only job is to watch the first stage. Did it complete the task? Did it use the right source? Did it make a claim it cannot prove? Did it write files but skip the rest? Did it do
12:06anything that needs human approval? The second agent is not there to be creative. It is there to be annoying in the best way possible. It is the person on your team who says, "Hey, this looks broken." That is how you get from toy automation to actual business automation. This also changes how I would compare Hermes and Open Claw. I would not compare them with vague questions like, "Which one is smarter?" That's irrelevant. That's the wrong question anyways. I would compare them on the same workflow. Give both tools the same job. Even the same repo, same file, same approval rules, same memory requirement, even the same failure case. Then watch what happens. Which one recovers better? Which one explains its failure better? Which one gives you cleaner logs? Which one respects permissions? And which one makes it easier to add a monitor? Which one turns a one-time fix into a reusable skill?
12:53That is the comparison that matters here. Because if you are building a real business with agents, you do not need a mascot. You need a worker you can trust. And trust is built by watching how it fails. This is why Hermes plus Codex story caught my full attention today. It points to the future that is better than most launch announcements do in a whole year. And the feature is not one giant magic agent that does everything perfectly. I mean, that sounds nice, right? But it is not how real systems actually work. The future is a team of smaller agents with clear jobs, clear permissions, and clear checks. One plans, one executes, one watches, one reviews, and then one reports, right? The human sets the direction and approves the risky stuff. That is the
13:38actual operating system of the work. So, what should you do next? If you were using Hermes, you need to try this pattern. Do not just ask Hermes to do random tasks. Give it one repeatable workflow and pair it with Codex as the watcher. Make Codex inspect the logs, the output, the state, and the final result. If you are using Open Claw to do the same thing, use Open Claw as the operator and Codex as the code workflow reviewer. Or use Open Claw to trigger a Codex subtask and then verify the output before anything goes public. And if you are new, do not start with 10 agents. Come on, what are you doing? Start with one. One worker and one watcher. That is enough. Because once you understand the pattern, everything changes, and then build it up from there. You really stop thinking like a prompt engineer, and you
14:25start thinking like an operator. And operators are the people who are going to win with this stuff. Hermes is not something Open Claw users should ignore. It's a serious comparison point. It is a threat in attention, and it is a competent in whole workflow ideas. And it is also proof that builders are ready for more than just chatbots. That is the part that I care about the most. The agent wars are not going to be won by who has the best demo. They're going to be won by who helps normal builders ship workflows that actually survive real life use cases. Hermes plus Codex is one example of the shift. And if Open Claw builders are smart, they will steal the pattern, test it, improve it, and make it boringly reliable. That's the whole game plan. If you need help setting up by your first AI agent, we do that
15:11inside of our community, and we have unlimited one-on-one calls to get your first AI agent up and running. So, if you join the community, you could book calls with our team, and we'll walk you through the steps, whether Hermes, open clock, quad code, we're here to help you. We also do seven live boot camp calls a week, so you can drop into one of those boot camp calls and get some information at get your question answered and we'd love to help you take the next step in your AI journey. We'll see you in the next one. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Hermes Desktop Tutorial: Install and Run Your First AI Agent
Hermes Desktop from Nous Research puts Hermes Agent in a chat-style app that took me about 15 minutes to install. I connected my OpenAI Codex plan with GPT-5.5 and ran three agents at once: a code repo review, a YouTube channel analysis, and an X account analysis. Then I set up a weekly cron job.
OpenClaw 5.2 Update: Cron Jobs, Context Safety, and Security Fixes
The OpenClaw 5.2 update is about reliability, not flashy features. Cron jobs run in their own lane, Codex and GPT-5 runs compact before they hit the context window, and approvals and secret checks are tighter. It also adds Discord voice guardrails and better model routing.
Hermes Agent Has 237 Real Use Cases Now
Hermes Agent has public receipts now. Here is what the 237 use cases actually prove for builders using OpenClaw, Codex, and persistent agents.
