Which AI agents I'm really using to build tech (Claude, Codex, OpenClaw & Hermes).
Most of my DMs are people asking which AI tools I’m using for software development, and what my thoughts are on OpenClaw and Hermes... Here you go.
Right now, my answer is pretty simple… but I’ll leave it towards. I still love Claude. It took me a while to adapt but when Claude Code came out last year, about a year ago, I immediately fell in love.
Claude is probably still my favorite model for writing, thinking, product strategy, and explaining complex ideas in a clean way. Generally speaking, but I do feel like the consistency and overall quality has dropped. Which is why this month, I switched back to ChatGPT (and Codex this month), after a long hiatus.
When it comes to coding, I have been back in Terminal using Codex again, disappointed in Claude Code in Terminal. Terminal, because I have not liked the newer Cursor update (Cursor 3). Cursor has been a big part of my workflow, but lately I have found myself wanting to get closer to the actual project again. Less UI. Less abstraction. More command line. More direct control. So, using Code and Codex via Terminal on my Macbook.
Claude Code is still extremely strong. I have used it a lot. It is very good at understanding a codebase, making changes, and moving fast when the project is clean. For a lot of tasks, Claude Code feels like having a very smart developer sitting next to you. But OpenAI’s Codex has been feeling better to me lately for troubleshooting, especially when something is broken, messy, or fragile.
The bigger shift for me is not just Claude versus OpenAI. It is not just Claude Code versus Codex. It is how these tools are being delivered. We are moving from chatbots to working agents… pushing the speed of development and troubleshooting.
I’m still jealous of those developers on Github who routinely push 100+ per day, or like OpenClaw’s Peter, 1,000+ pushes per day. My most intense days are about 50 pushes per day. To speed things up and to tackle more complex issues, I have been testing OpenClaw through Terminal and Discord. OpenClaw is a little fragile in my opinion. I haven’t loved it.
I have also been testing Hermes through Terminal and Telegram, and that has been really great at problem solving. Right now, for complex projects, my preferred workflow is Hermes plus Codex, especially through Telegram.
That may sound strange, but it actually makes a lot of sense. A lot of real coding work does not happen in one clean session. You do not just say, “build the app,” and it magically ships. Real projects break. Packages conflict. Environments fail. Auth gets weird. Stripe needs another pass. Supabase rules need to be checked. The build fails. The mobile wrapper works locally but breaks when packaged.
This is where most vibe coders get stuck. They can get to 70 percent. Sometimes 85 percent. But that last part is where the real work starts. The finish line is not the fun part. The finish line is testing, debugging, packaging, submitting, fixing edge cases, dealing with Apple, handling permissions, cleaning up errors, and making sure the thing actually works for a normal user.
That is why I am paying close attention to agents. Because the winning workflow is not going to be one magic model. It is going to be the best system for getting work done across multiple steps.
Claude is great for thinking. Claude Code is great for moving through a codebase. Codex is feeling strong again for Terminal based troubleshooting and execution. Hermes is interesting because I can work with it through Telegram and keep projects moving in a more natural way. “Self improving” they claim, definitely expansive. OpenClaw is interesting because it points toward a future where agents are not trapped inside one app…
So my current answer is simple. For complex coding and troubleshooting, I am leaning toward Hermes plus Codex. Not because it is perfect. None of this is perfect. But because it gets me closer to the actual work.
And that is the whole point. I do not care which company wins the AI branding war. I care which tool helps me ship working software. That is where the market is going. The next wave is not “who writes the best prompt?” The next wave is who can actually get the project finished.
Inspired by my TikTok video:
Tiktok failed to load.Enable 3rd party cookies or use another browser
CTO Dave
David Levine Bramante
Fractional CTO
david@codexdeus.com
codexdeus.com
text/whatsapp. (310) 906-5459