claude opus 4.8 doesnt even come close in terms of understanding how systems actually work or bothering to actually read existing code beyond the most cursory glance. the united states empire is truly over, as are these “premium” “flagship” closed weight Western models. Every so called mainstream “benchmark” publication is a bold faced lie where each company pays off the creators of each major benchmark publisher to tip the scales in their favor. Even when playing dirty, however, openai and anthropic still can’t win.

it has anything to do with my external calibration stuff so no

You are right. People do not care about agorism. But, Tailscale is not something that should be recommended. There are actual real self-hosted options like Netbird and Nebula.

just the nature of them being quite old models without proper tool calling functionality. What actually DID help was setting up middleware and custom python servers/clients with proper json mapping to enable the proper tools to be selected. so, literally zero model tuning required in the end.

unfortunately, i did not notice much of a difference with model tuning. it took a pretty decent chunk of time. For my most powerful pc, which is what I run most models (the lower end machines with worse gpus run embedded text models) I got a fairly powerful machine with a single 4090. I have had better luck just downloading differently tuned variants of the same model from others

lora, yes. mostly custom scripts downloaded off of hugging face to automatically handle a lot of complicated stuff I’m not totally sure of how it actually works under the hood to be honest

you can try to ask the agent to talk like a pirate for example

hermes has some built in agent orchestration layer which seems cool on paper. never tried it. other small nice things that are unique to hermes which other agents really don’t have, which I have actually tried, include: switchable agent personalities, pretty decent thread suspension mechanism, decent webhook subscriptions, and human delay mode. The biggest thing, at least in my opinion, is certainly: Self-improving skills with patching - with an entire slew of caveats… In my opinion, this is useful but I strongly recommend using a manual review process. Otherwise, the agent has the potential to “teach itself wrong”. Human review.

working on it this very evening. rootless podman support for each database and each external worker container. My openrc implementation is already working! implementing a systemd service implementation with timers instead of fcron as well as was planned. you want quadlets? quadlets will be supported too! and all of it will integrate with the new healthcheck and self-repair system

filesystem access is something you can choose to give it, but the security model by default doesnt allow this. you can enable certain flags within the database that could allow local filesystem access to use tools specifically designed for that but that is not anything the agent can do by itself. This is an example of a “developer tool” and its not something I use at all outside of a single tenant instance on a raspberry pi 5. Sandbox and worker containers are preferred. The agent’s memories, routines, settings exist within a database, not on a file system. a basic internal shell can be “emulated”, but for advanced tasks, that’s what the sandbox or external workers are for. That being said, feel free to use the mt admin script to automatically create users anyway for multi agent setups. it takes care of linger, adding the user to a docker group, and so on. Permission system is incredibly robust AND customizable. The default permissions are generally sane though: Tool calls need explicit approval by default and approval prompts pause other agent activity in the agent loop in the conversation thread. Up to you if you want to pass “always approve” to specific tools

Yep, I got harnesses for the actual multi-tenant setup that support pre-seeded values and memory workspace data as well as testing harnesses to automate a lot of the developer testing. Additionally, I’m writing a chaos harness as well for chaos engineering tests in order to test the new self-healing feature

I would suggest two models: a local gemma 4 26b variant or a cydonia variant. This one specific custom gemma variant I have been using does a very good job at providing vivid details, while not being overly verbose. I love Deepseek 4 but I can assure you I agree with you 100% in that I don’t think it can write “like a human”. I’ve also got some custom versions of GLM4.7 that support very high memory context that run well on consumer hardware and also seem to have very good writing styles. I’ve also used various GLM 5.1 instances in remote TEEs are also good at writing, world building, etc. There’s plenty of other suggestions I’m sure others will give you, such as Qwen variants.

However, model suggestions out of the way:

Writing styles are always going to be tough to get right, regardless of the model.

Personally, I’ve adjusted dozens of sampler values, written middlewares, llama-server scripts and configuration loading mechanisms, openai api compatible HTTP proxies, and even a python3 API for accessing context information and being able to switch models on the fly. I’ve even created a local model benchmark performance script.

But besides running some scripts which others have made to tune a model specifically with specific input parameters, not really. Honestly, I have a lot to learn.