Taste Agent

This project is testing out a light non-consensus thesis I have: that in a multi-agent future, we’ll all want an agent dedicated 100% to understanding ‘us’ and injecting that into every workflow with other agents.

Other agents will specialize in tasks, verticals, projects, types of work, etc. based on their harness and training. And they’ll have access to our preferences, memory, data, etc. in many cases. But the thing that differentiates ‘me’ from anyone else, more than anything else, isn’t actually captured well in those discrete files — it’s the taste, judgement, patterns and nuanced day to day things that can only be inferred from a large amount of experience and is best captured in weights, not in files.

This project experiments on that thesis that taste should live in one deep node that other agents consult, the way you’d ask the colleague whose judgment you trust rather than hand everyone a style guide. It started as a note to myself: “make a model that is purely embodying my taste - find right tech or way to encode this (maybe not text) and store it. Use as a way to learn a level deeper. And make a partner everywhere.”

I didn’t start with a model right off the bat. For now it’s a thin agent storing a record of my judgments: what I approved, what I changed, what I rejected, and why. It manages, updates, prunes and accesses that record and interfaces with other agents in my workflows to see if having that dedicated expertise improves outcomes and alignment to ‘my taste.’

The first experiment captured judgment moments from my own sessions with coding agents and used a small local model to sort out the ones that mattered. The lesson was about order: capture with nothing using it stalled, so I flipped it and started serving judgment first. Now, before an agent shows me a draft or a plan, it asks the taste node, which critiques the work against rules I’ve written in my own words and cites the rule it’s applying. When I override it, that becomes the next example to learn from. It also has a removal rule built in: if it goes unused for two weeks, or I overrule it every time, the agents stop asking it.

In time, if this succeeds, I’ll move to a LoRA adaptor or even post-training my own open weight model to really encode this more deeply. For now, it’s active and changing week to week. The open question is whether a few hundred of one person’s judgments can teach a model something a general one misses.

Update: October 2026

A month in, agents consult the taste node about six times a day. Between October 2 and 27 they asked it 160 times, mostly to critique a draft or a plan before showing it to me, and sometimes to ask whether I’d approve a choice before bringing it to me. 85% of the critiques came back “revise”, each with three or four concerns tied to a rule I wrote.

It changes what reaches me. The first draft of my Design Harness writeup opened with “Every design session with an AI started from zero.” The node flagged that as a problem every reader already knows, with the surprising part, a critic that only ever sees screenshots, saved for last. The revised draft opens on the critic, and the process paragraph is half as long. Elsewhere it caught a buried lede, a document count that didn’t add up, and figures in an email that didn’t reconcile, all before I saw them.

Most of what it’s asked about isn’t voice or design. I built it imagining those, but most consults are about plans and proposals, and the rules it leans on hardest are about how to work: lead with where things stand, surface the decision instead of burying it, prove a claim rather than assert it. In agent workflows, most of what I’d call my taste turns out to be judgment about how we work together.

The logs also hold the first real test of the multi-agent bet. A third of all consults came from Antigravity, Google’s Gemini-based coding agent. It has the node in its tool list but none of the instructions I gave my Claude agents about when to use it. Working only from the server’s one-paragraph description of itself, it started checking its own replies with the node before sending them. That’s the behavior the project is betting on: an agent built on a different model, by a different company, consulting the thing that holds my judgment with nothing more than that description to go on.

The weak spots are just as clear. Asked “would Danny approve this?”, it says yes 90% of the time, usually to whichever option the agent listed first, which is almost always the agent’s own pick. Critiques almost never come back clean, so a “revise” means less than it looks. And it’s been wrong in ways that aren’t me: twice it passed an email I then cut shorter and made less certain, and on a piece of Sous copy it told me to trim a story I’d told long on purpose, because the story only works if you paint the future you’re arguing against.

So it works as a reviewer. Whether it judges the way I would is still unproven, and that’s what the learning half is for. I haven’t started it yet. The bar is deliberately high, 100 recorded reactions across at least two kinds of work before any training run, and so far it has 20. The agents I was working with logged all of them: some are my rewrite of a draft, some are the option I picked, and none is me telling the node directly that it got something right or wrong.