Kenneth Ballenegger

VC / Contrarian / Dog Dad

[Kenneth Ballenegger @kob]

@pgllmt Here's the overall architecture of it, though it's evolved a bit since I wrote this.

Will be coming up with a productized version of this for anyone to use with their AI agents as soon as it's polished enough. Still a work in progress.

https://t.co/hLSuY8ZRhG

[Kenneth Ballenegger @kob]

Will take angel investment in the form of Mac Studios at retail price and DGX sparks.

[Kenneth Ballenegger @kob]

@stanine I wonder why AI talks like this. Surely that's not a big pattern in the training data of how humans actually talk

[Kenneth Ballenegger @kob]

Why does AI love the word "wedge?"

And since when can a stuck / frozen process be referred to as "wedged?" I've never heard the word used that way.

[Kenneth Ballenegger @kob]

@HermesWatcher Capable generalist; with its capabilities defined in lazy loaded skills.

[Kenneth Ballenegger @kob]

I don't want Hermes running on my main machine, I run it on a dedicate machine that doesn't have access to my most important accounts.

I'd rather not have that code present on my main machine at all. For security / cleanliness reasons.

When I tried to install Hermes Desktop out required downloading and running Hermes locally before I could get to the screen to connect to a remote Hermes.

[Kenneth Ballenegger @kob]

@multiplanet1 AI still doesn't know how to count.

[Kenneth Ballenegger @kob]

@MarcosHernanz I have the opposite insurrection, never install any dependencies that aren't absolutely necessary and if so only with human approval.

[Kenneth Ballenegger @kob]

@haridigresses @rjonesy Any resources showing how it is being used to such effect?

[Kenneth Ballenegger @kob]

Interesting, that's very detailed reasoning.

Since Claude code's system prompt has leaked and since Codex's is open source (and poor), have you played with replacing Codex's with Claude?

Good point regarding queue — I haven't been using it with any of them but it's worth trying.

One thing that makes Claude better for me is it has much better non-interactive (claude -p) behavior then Codex including the ability to steer a non-interactive session.

I'm currently trying to get every model into Claude instead of using their respective harnesses.

[Kenneth Ballenegger @kob]

I switched to Hermes right when it came out too and feel like I only use the featureset available then. All limitations at the time I've built custom solutions too. Some of which Hermes has evolved to fix too— but now it's too much work to reevaluate from scratch rather than half my custom stack on top of Hermes.

[Teknium 🪽 @Teknium]

🫡🤗🤗

[Kenneth Ballenegger @kob]

Damn, I built my own local version of this feature but seems I've been sherlocked.

Might need to plug my app into the official API.

[Nous Research @NousResearch]

Hermes Agent now has voice activation.

Say the wake word and Hermes opens a new session and listens for your command, hands-free in the CLI, TUI, or desktop app. Detection is local and off by default.

https://t.co/V5DwesUdod

[Kenneth Ballenegger @kob]

Or find a travel agent who can book you a Hyatt Privé rate. You will get breakfast included, an upgrade, 4pm checkout, $100 dining credit, and sometimes 25-33% discount. No status required.

I can book this rate for anyone, consider this an open invite to DM me for help.

[Kevin @Kevincreates77]

A couple booked a Park Hyatt in Tokyo for their 10th anniversary.

They paid the standard rate. $620/night. 4 nights. $2,480 total. Standard room. No upgrades. No lounge access. No breakfast included. They checked in, got their key, and went to a room on the 7th floor with a city view partially blocked by a neighboring building.

In the elevator, they met another couple staying the same 4 nights at the same hotel. Same dates. Same anniversary trip.

That couple was staying in a corner suite on the 22nd floor with a panoramic view of the Imperial Palace. Free breakfast every morning in the club lounge. 4 PM late checkout on the last day. Resort fees waived. Total paid: $1,800.

Same hotel. Same dates. 4 nights. $680 less. A suite instead of a standard room. Free breakfast instead of $65/person downstairs. Late checkout instead of the standard 11 AM rush.

The difference wasn't a special deal or a corporate rate. The difference was 9 features inside the World of Hyatt loyalty program that one couple used and the other didn't know existed.

The elevator couple told them everything over dinner that night.

Here's the full playbook 🧵

[Kenneth Ballenegger @kob]

Not having a "personality" doesn't mean not having a system prompt defining how it behaves. It should of course have that.

But the common abstraction of giving your agents anthropomorphic personalities (e.g. the reviewer, the coder, the researcher) I feel is not a good abstraction.

Instead my agent has its small generic system prompt defining how I'd like it to behave; and a bunch of skills loaded when relevant (how to code, how to review, how to research) etc. The agent uses the same memory system across all runs, loads relevant memories.

Sub agents is just parallelism. The agent can run parallel background jobs; those can have fresh context to run cleaner. We don't need the fake "My architect agent is managing 12 junior dev agents" story when we can just say "claude code is running 12 parallel jobs."

[Kenneth Ballenegger @kob]

I’m skeptical of how the AI industry frames agents, sub-agents, and agent personalities.

There's no such thing. Agents are just LLM calls with system prompts and tool calls and memories. Sub-agents are just threads or parallelism with a fresh context window.

I'm not convinced these abstractions add any value. One agent that understands its capabilities and parallelism is just as good.

What am I missing?

[Kenneth Ballenegger @kob]

@krunkosaurus Abstractions can be an extremely powerful tool when used right. OOP made sense (though I was much more a fan of functional programming) and was a useful abstraction.

My hot take is I feel like agent personalities is not a useful abstraction; it's counterproductive and wasteful.

[Kenneth Ballenegger @kob]

@PascalPixel I run local models to process all my email privately. Qwen 35b is more than capable of classifying my inbox (this has been tested against frontier) and automatically processing all of it

[Kenneth Ballenegger @kob]

That's on point— but I'd go further. Nobody will be reviewing code much longer, so we need to invent a way to tie a new 'git blame' to the live app and unexpected changes that occurred, and tie that back to the commits and prompts that triggered them

[Konsti Wohlwend @konstiwohlwend]

tbh i don't just want a GitHub clone, i want GitHub reinvented.

i don't want to review source code, i want to review the chat history that produced it. git blame should show me the prompt behind a change.

developers think in prompts now, not code. why aren't they committed to source control?

right now we're doing the equivalent of sharing binaries and asking our team to decompile them. there must be a better way.