KAVI

AI Systems Engineer | Agentic Infra, Orchestration, Context & Model Routing | AI + Markets | Personal research, not financial advice

[KAVI @0xKAVI01]

I build AI systems that have to survive outside the demo.

Agentic infrastructure, orchestration, context, model routing, tooling — plus the occasional market experiment.

I’ll share what I build, what breaks, and what actually works.

[KAVI @0xKAVI01]

I barely need to explain anything. The way I talk to @Muse understands basic language, asks questions to clarify what I actually want, and delegates to the right agentic infrastructure. Each harness has its own infrastructure I also built. For context,Muse built one canonical Agent native shared repo and shared Obsidian context. I’ve been offered by Muse to make a dashboard. I don’t have time for that. I am a phone guy; I run the majority of the operations with an iPhone. I’ve been in the agentic era since 2024.

[KAVI @0xKAVI01]

Two months ago I started Building Hermes From Scratch because I wanted a persistent AI operator, not another browser tab with a chatbot.

Since then, my AI stack has become a lot more interesting.

I'm building around Nous Research's Hermes Agent. Muse now helps manage my Hermes environment, while I use Codex, Claude Code and ChatGPT for different kinds of work.

Here's the part I find fascinating: I don't think the future is one AI super-app that makes every other tool disappear.

I think it'll look more like a small studio. One system coordinates. Another builds. Another researches. You decide the direction and step in when your judgment matters.

The industry is obsessed with which model wins the next benchmark. Fair enough. I follow those releases too.

But working with AI agents changes what you pay attention to. It's less about which chatbot gives the prettiest answer and more about how much you can actually get done with the tools you already have.

I started the original Hermes series to explore architecture. I'm bringing it back to share the experiments, model comparisons, surprising tradeoffs and things I'd build differently.

I'm not chasing the fantasy of a company with no humans.

I'm chasing the freedom to stop pushing every button myself.

The model race is exciting. The race to make these tools useful together might be even bigger.

What AI tool has genuinely changed how you work?

[KAVI @0xKAVI01]

@NousResearch @Teknium Please tell me hermes app will able tomuse your iphone ?

[KAVI @0xKAVI01]

@aksenHQ @Muse Faceless yt/Shopify dropshipping/local business consulting/ ai interference setups (local/cloud)/lead selling (b2b)/website design (agent native)/personal assistant builds/Etsy POD or digital products. I have a couple more projects, but personally I can’t keep up with context.

[KAVI @0xKAVI01]

Last update: I gave the keys to @Muse to run my Codex, Claude Code, and Hermes on the same VPS, layered by different sandboxes. Layered by different sandboxes . 32gb ram 4vcpu 200gb backed by 2tb cloud. Total cost$102/ month. Running 9 different business models generating $7,800/month. Finally, it feels like I built the harness I wanted to.

[KAVI @0xKAVI01]

@HermesWatcher Cua/advanced delegation

[KAVI @0xKAVI01]

@Pricerrors Thats how you wash money

[Akshay 🚀 @akshay_pachaar]

Finally, an OpenRouter for agent harnesses!

(including System One by Jev)

Devs just open-sourced a plug-and-play infrastructure layer that lets you run any harness under a single interface, like:

- Codex
- Hermes
- Claude code
- DeepSeek Harness
- System One, powered by Jev
- And 9 more agent harnesses

This means you can bring Jev into the same product that already uses Codex, Claude Code, or another supported harness, without writing another implementation for sessions, streaming, files, cancellation, and failure handling.

Here's the repo: https://t.co/scJwb2xUyM

(don't forget to star it ⭐ )

The harnesses run locally, and the Unified Harness Protocol (UHP) defines the common task interface with an OpenAI Responses-compatible API.

If you want to dive deeper, my recent article explains why model routing is not the same as harness routing, and what it takes to support multiple harnesses.

It also covers UHP, the full local setup, a working API call, and how sessions and files work.

Read it below.

[KAVI @0xKAVI01]

Vercel released durable-harness: coding agents can recover after crashes without blindly repeating completed work. Honestly, this excites me more than another benchmark win. Building while your laptop is closed should be normal.

[KAVI @0xKAVI01]

OpenAI's GPT-6 Luna Decisions API can now handle up to 400 bounded questions per request. I build multi-model agent workflows, so this interests me more than another leaderboard. I haven't tested the endpoint yet. My first test: routing accuracy vs cost.

[KAVI @0xKAVI01]

Wrong hands always

[KAVI @0xKAVI01]

@harden_369741 Check out Hermes MoA. But i also suggest Jev layer for decision mechanisms.

[KAVI @0xKAVI01]

Vercel Labs' fx is a fascinating challenge to bloated AI coding agents.

It's an open-source CLI written in Zig. Instead of replacing your terminal, it preserves Unix-style scrollback. It supports ChatGPT/Codex and Grok subscriptions, MCP, subagents, and embedding through libfx.

I use Codex and Claude Code in my engineering workflow. What interests me here isn't another model benchmark. It's whether the agent can be composable, scriptable, and get out of the developer's way.

I haven't tested fx yet. I'd compare it against Codex and Claude Code on the same repo-edit task, then judge the actual developer experience, not just the generated code.

Repo: https://t.co/2Wlqpo5ZpS

Would you switch coding agents for a better terminal workflow, even if the underlying model stayed the same?

[KAVI @0xKAVI01]

@harden_369741 Agreed. With one difference . Precise execution with reliability is the product now imo. Workflow is just a skin. You don’t even need intelligent models for running repetitive work.

[KAVI @0xKAVI01]

@evgheni_D Mcp is most important thing lately

[Tibo @tibo_maker]

https://t.co/DfsTmI4zPE

[KAVI @0xKAVI01]

@AlexChhk They just approved my premium after 2 weeks. Lmao

[KAVI @0xKAVI01]

Hot take: most 'AI businesses' are chatbots attached to forms. Give me the boring agent that answers missed calls, books appointments, and chases unpaid quotes. That's easier to sell than another dashboard. What's the most absurd manual workflow in your industry?

[KAVI @0xKAVI01]

OpenAI models are coming to Google Cloud. Read that again: Google's cloud will sell a rival's frontier AI. The model war is becoming a distribution war. The company that owns the enterprise checkout may matter as much as the one that trains the model.

[KAVI @0xKAVI01]

Claude Duet just joined Muse and Dots. Every AI lab wants an agent running your day in the background. The winner won't be the smartest chatbot. It'll be the one you trust with your calendar, your work, and its own mistakes. Who gets your keys?