Thinking about Hermes Agent
Weblog notes from June 16, 2026 that were compiled and shared
I have been using Anthropic’s Claude Opus 4.8 model within the Hermes Agent on my MacBook Air for the last couple of days. Part of that time I used a couple of the free models available in the Hermes Portal, and the difference between Opus 4.8 and the free models is night and day when it comes to coding and speed. Seriously, the amount of things that Opus 4.8 will just trigger and do to push a coding project forward is seriously interesting. Everything defaults to TDD, and the model plans and takes action at an amazing speed. A lot of the decisions and actions it takes were not what I expected during the development process. Some of the decision-making seems to be just autopilot compared to me having to provide direction on how to engage in the development. The Hermes Agent paired with Opus 4.8 seems to just code within a pattern and framework that is well established. This is something that was new to me, and the experience made me sort of question how the orchestration is happening. A lot of the common or accepted development methodology seems to just be built into what the Hermes Agent is structuring for code development.
Maybe what I’m trying to say here is that the development path forward using these types of agents is going to end up making code development abstracted going forward. I’m watching the development on GitHub and looking at what is being made, but the Hermes Agent needs very little help from me or input on the development choices being made. I’m more or less along for the ride and providing a few instructions, but I’m not the primary driver as the project continues in a very session-focused series of development efforts. Sure, we built a plan and worked on what would end up being worked on as a part of the process. The interesting part of the process was just how many tokens were used in the planning parts of the process. The overall spend on the Opus 4.8 model had these spikes for planning and then a more regular ongoing cost for the session development work. Opus 4.8 is amazing to work with. It is expensive to run, and I ended up watching the cost in tokens on a minute or hourly basis. During the start of this process, I decided that I was willing to pay for a couple of days of tokens to get a feel for how the development process with Hermes Agent and Opus 4.8 would work.
What exactly have I been working on during the last few days? I have worked on the Knowledge Reduce GitHub code that I had previously run about 10 different Manus sessions refining and working on, which has been published to GitHub over the last couple of years. I decided to take my most advanced work as a starting point and give the Hermes Agent my paper on the subject to refactor and produce a complete code package for that effort [1]. During the course of working with Manus, I had sessions that ran for hours on end and produced some interesting results, but none of that ever really turned the corner [2]. This current turn of the wheel with Opus 4.8 did manage to turn the corner, and the code that has been developed actually works now and can be verified by runs against my test data set, which is very interesting [3]. Assuming that you are willing to spend the token cost, the shift in what is now possible is very real, but it is very expensive to execute. The economics of token usage is now a very real factor in deciding what projects to work on and what coding efforts should be tackled based on cost.
Footnotes:
[1] https://github.com/nelslindahlx/KnowledgeReduce/tree/main/Paper_Preprint
[2] https://github.com/nelslindahlx/KnowledgeReduce
[3] https://github.com/nelslindahlx/Hermes-Output/tree/main/knowledgereduce