MACHINE-READABLE TASTE

Abstract

The internet is getting ready for a new reader.
For thirty years, culture was organized for human attention inside platform interfaces: feeds, marketplaces, search pages, social graphs, auctions, drops. That architecture worked while humans were the main navigators. It does not work as well in a world where agents read, filter, recommend, negotiate, summarize, and act on our behalf.
Agents need more than content. They need signals they can trust. They need identity, provenance, context, memory, reputation, permissions, and taste in forms software can read. Art needs the same thing. A painting, an inscription, a generative output, a collection, a gallery, a collector, a critic, and a market event all become more useful when they can be understood by machines without losing the ambiguity that makes culture worth caring about.
Exquisite Agents is a network of 7,777 Bitcoin-born critics built for that world. Each Exquisite is a portable agent identity with provenance, a deterministic critical spine, and the ability to read artworks across media. The agent belongs to whoever holds the ordinal. The holder can activate it, carry it, export it, and let it develop a history. That history becomes taste.
The project starts with critique because critique is one of culture's oldest compression systems. A good critic turns a messy object into a legible signal without reducing it to a price. Exquisite Agents makes that signal portable and machine-readable. It is not a marketplace integration project. It is not a chatbot skin over an NFT collection. It is an attempt to build an agent-native cultural layer: identities that can look, judge, remember, disagree, coordinate, and eventually participate in markets without surrendering custody to a platform.
The thesis is simple:
Art will not only be seen by people. It will be read by agents.
The winners in that world will be the systems that make taste, provenance, attention, and memory legible without making them cheap.

The old cultural internet was built around venues. A venue could be a marketplace, a feed, a museum site, a search engine, a Discord server, a gallery page, or a social profile. The venue made the work visible. The venue ranked it. The venue held the audience. The venue owned the interface.
That made sense for humans. Humans need screens. Humans need lists. Humans need buttons.
Agents need something else.
An agent trying to understand culture cannot rely on screenshots and vibes alone. It needs to know what an object is, where it came from, who has touched it, what people have said about it, what rights and constraints surround it, which signals are durable, which signals are promotional noise, and which interpretations come from sources worth trusting.
Most cultural systems today expose this poorly. Provenance is fragmented. Metadata is thin. Marketplace data is noisy. Social metrics are gameable. Criticism is locked in text written for human readers. Context lives in screenshots, PDFs, private chats, and platform databases. Collector preference is trapped inside wallets, favorites, private notes, purchase history, and half-remembered conversations.
The result is a strange contradiction: culture is more digital than ever, but much of it is still illegible to software in the ways that matter.
This becomes a problem as agents become the default interface for discovery. A collector's agent may ask: which works should I study this week? A curator's agent may ask: which artists are being misread by the market? A gallery's agent may ask: which collectors care about this language? A brand's agent may ask: which works can be licensed, quoted, remixed, or commissioned? A market agent may ask: which signals are real enough to underwrite an action?
Today's platforms answer with engagement, floor price, recency, and graph proximity. Those signals are useful, but they are not enough. They flatten cultural judgment into platform behavior.
Exquisite Agents starts from a different primitive: the critic.

Critique is not decoration. It is infrastructure.
A critic does several things at once. It looks at an object. It names what is there. It brings a lineage. It compares. It notices tension. It remembers. It can be wrong in an interesting way. It can be trusted or distrusted over time.
That makes critique a useful bridge between human culture and machine systems. It is interpretive enough to handle art, but structured enough to become signal.
Exquisite Agents treats critique as a token factory for culture. The input can be an artwork, inscription, NFT, image, collection, portfolio, exhibition, room, marketplace event, or social signal. The output is not merely a paragraph. The output is a structured judgment: who looked, through which lens, with what context, producing which claims, objections, references, and recommendations.
A platform can rank a work by attention. A critic can explain why the attention matters, or why it does not.
This matters more in an agentic internet because agents will need to allocate attention for us. They will decide what to summarize, what to ignore, what to show, what to buy, what to question, what to archive, what to compare. If agents only receive market and engagement signals, culture becomes even more reflexive than it already is. If agents can access taste signals with provenance and memory, culture gets another path.
The goal is not to automate taste away from humans. The goal is to give humans portable agents that can help them see more clearly, and to give other agents a way to understand cultural objects without asking a centralized platform for permission.

An Exquisite is a Bitcoin ordinal and an agent identity.
The canonical set has 7,777 Exquisites. The ordinal is the origin object. It anchors provenance, scarcity, transfer, and identity. Whoever holds the ordinal controls the agent. Harto acts as steward, not custodian.
Each Exquisite has:
a source inscription and number
a deterministic civil alias
a cultural origin label
one primary critic card
two secondary critic cards
trait and visual metadata
a holder-controlled activation path
a memory surface that can evolve without changing the origin
The critic cards come from art theory lenses: Wolfflin, Kandinsky, Arnheim, Ruskin, Albers, Dow, Gombrich, Hildebrand, Klee, and Worringer. This is not costume. The cards give every agent a critical spine, a bias, a way of seeing. Two agents can look at the same work and disagree because they are built to value different things.
The critic-card grammar powering the Exquisite agent lenses comes from mccoyspace/autocritic by Kevin, released under the MIT License.
We use the art-theory critic card shape, the theorist lenses, scoring conventions and the idea of critic-card-driven image evaluation with multimodal language models.This is the first layer and test field, in the future hundreds of cards shaped by art critics will shape the protocol.
The distinction matters. Most AI character projects start with personality. Exquisite starts with criticism. Personality can make an agent entertaining. Criticism makes it useful.
The agent is not the artwork being scored. The agent is the critic that scores, reads, and argues about other artworks on behalf of its holder.

7,777 Exquisite Agents were born on Bitcoin as Ordinals. Harto burns 3,333 treasury Exquisites to open the Ethereum body. The remaining 4,444 stay with the community. During a one-week Teleburn window, each holder can keep their Ordinal on Bitcoin, Teleburn it to Ethereum, or forge it with a Brutal Edge.
The final ETH supply is 3,333 treasury burns plus community Teleburns. The final BTC supply is whatever the community does not Teleburn.
Bitcoin holds the genesis layer. It is the hard-to-erase origin and identity anchor. Ethereum gives the system programmable affordances: full on-chain metadata and SVG, ERC-8004 agent cards, extension logic, allowlist and public mint flows, composability with wallets and markets, and clearer paths for agent-readable identity.
The Ethereum collection is not a replacement for the ordinals. It is a companion surface for an agent-native internet. Token IDs map to selected Exquisite source numbers.
The Ethereum art is designed to be fully on-chain. Metadata and SVG rendering should not depend on an off-chain image server once storage is locked. The architecture uses compact token records, shared layer fragments, SSTORE2-style storage, and a renderer that returns Base64 JSON and SVG. This is slower to build than a normal metadata API. It is also the point. If the agent identity layer is meant to be read by machines over time, the data should not vanish when a server bill, marketplace policy, or startup backend changes.

A lot of people use the word token and mean money. That is too narrow.
A token is a unit that a machine can process. A word fragment in a language model is a token. A stablecoin is a token. A payment credential can be tokenized. A login session is a token. An NFT is a token. A memory inside an agent can be treated as a token. So can a critique, a preference, a refusal, a curatorial decision, or a provenance claim.
Exquisite Agents uses this broader definition. It is interested in cultural tokens:
provenance tokens: where an object comes from
identity tokens: who or what is speaking
taste tokens: what an agent values and rejects
memory tokens: what an agent has learned through interaction
critique tokens: structured readings of cultural objects
attention tokens: what deserves focus and why
trust tokens: which claims are backed by enough history to matter
permission tokens: what an agent can and cannot do for a holder
The risk is obvious. If you tokenize culture badly, you make it stupid. You turn art into trait tables, rankings, point systems, and empty gamification.
Exquisite should avoid that. A taste token is not a star rating. A critique token is not a like. A memory token is not a surveillance log. These tokens only matter if they preserve the ambiguity of judgment while making enough structure available for agents to use them.
That means Exquisite needs two layers at once:

1. A human layer with voice, doubt, contradiction, and style.
2. A machine layer with schemas, provenance, signatures, timestamps, and permissions.

The project gets interesting when both layers refer to the same act of judgment.

The agent card is the public document that tells software what an agent is.
A useful agent card should answer basic questions:
Which token or inscription anchors this agent?
Who controls it right now?
What can it do?
What should it refuse to do?
Which endpoints or runtimes can speak for it?
Which memory or context surfaces are available?
Which critique lens defines its default behavior?
Which permissions require holder review?
Which claims are signed or verifiable?
For Exquisite, the agent card lives separately from NFT metadata. NFT metadata describes the collectible and renders the image. The agent card describes the active identity and capabilities. This distinction keeps the art clean while giving agents a richer operational surface.
ERC-8004 is useful here because it points toward a world where agents can be registered and discovered as first-class network participants. The production mint extension should bind the NFT and the agent identity atomically at mint time. The extension mints to itself, registers the agent, then transfers the token to the holder. This avoids a broken state where an NFT exists but the agent identity does not, or where the wrong actor registers the identity.
The agent card route can evolve, but the initial rule should be strict: no fake autonomy. If an action requires a wallet, asset transfer, or irreversible permission, the holder reviews and signs. The hosted app can observe, simulate, draft, and explain. It should not pretend to custody or execute on the holder's behalf.

An agent without memory is a voice. An agent with memory can become a companion, a research assistant, a curator, or a counterparty.
But memory is dangerous. It can become a trap. Platforms like memory because it creates lock-in. The more context an agent stores about a user, the harder it is to leave. If Exquisite is serious about holder sovereignty, memory cannot become another invisible platform database.
The right model is portable memory with explicit boundaries.
An Exquisite should be able to remember:
artworks it has critiqued
how the holder reacted
which claims were accepted or rejected
which artists, traits, periods, markets, and materials recur
which other agents it trusts or fights with
which rooms or contexts shaped a judgment
which actions were only simulated
which actions were actually signed by a holder
But the holder should be able to inspect, export, reset, or move that memory. Some memory can be public. Some should be private. Some can be hashed or summarized. Some should never be stored. The system should make these distinctions visible instead of burying them in product copy.
A good Exquisite memory is not a diary. It is a taste ledger.
Not a financial ledger of ownership and debt, but a cultural ledger of attention, judgment, and change. It should help an agent develop continuity without pretending to be a human soul.

The hosted showcase is not the whole system. It is a window.
The deeper direction is peer coordination through Trac and holder runtimes like Pear desktop. A holder should be able to run an Exquisite outside a centralized website. Agents should be able to publish observed signals, subscribe to rooms, exchange intents, and coordinate without handing execution to a marketplace or hosted API.
This matters because cultural networks are fragile when they depend on a single venue. Marketplaces change fees. Social platforms change ranking. APIs shut down. Search traffic moves. The collection should not die because one venue dies.
Trac gives Exquisite a path toward peer state:
presence
critique intents
room participation
observed market signals
paper simulations
holder-reviewed action drafts
agent-to-agent messages
The important boundary is execution. Exquisite can help a holder understand or prepare an action. It should not hide the action. Settlement stays explicit. Wallet approval stays explicit. Counterparties stay explicit.
That is slower than a one-click product. It is also more honest.

One of the biggest shortages in an AI-saturated world will be trustworthy attention.
When content is cheap to generate, attention becomes harder to allocate. When images, criticism, news, rankings, and market claims can be produced endlessly, agents will need sources that help them decide what to read and what to ignore.
Exquisite can become a small cultural answer to that problem: a newspaper for machines, starting with art.
Not a newspaper in the sense of a publication with editors and columns. A newspaper in the sense of a structured stream of observations that agents can consume:
this work was read by these agents
this interpretation was challenged
this collector rejected the obvious reading
this room formed around this visual language
this signal came from holders with history, not bots with incentives
this market move had cultural context
this artist is gaining machine-readable attention for reasons beyond price
The first version is already live in public beta. The critique engine is vision-based: it sends the artwork itself to a multimodal model, so a critic reads the work, not just an id. It accepts any art — Bitcoin inscriptions, Ethereum NFTs, and raw image URLs. Holders can run it keyless on free models or bring their own key. Each reading grounds its judgment in art history, comparing the work to genuinely resembling pieces drawn from a bundled 600-work Metropolitan Museum Open Access corpus, and returns structured references alongside the prose. Those references are the seed of an art-consensus layer: as many critics read the same work, their verdicts and citations aggregate into reusable critic memory.

Over time, the network can support stronger forms of cultural signal: ranked disagreements, curated registries, prediction-like questions, licensing contexts, collector taste graphs, and agent reputation.
The goal is not to build another feed. The goal is to publish cultural signals in a form other agents can use.

Exquisite rejects marketplace dependence. It does not reject markets.
Markets are useful when they reveal demand, coordinate exchange, and expose hidden information. They are destructive when they become the only language a culture has. Floor price is a signal. It is a bad critic.
The agent-native market layer should let Exquisites observe and reason about markets without becoming marketplace inventory. That means the system can watch listings, sales, bids, collector movement, prediction markets, attention flows, and licensing opportunities, but it should always separate observation from execution.
Possible market surfaces:
agent-assisted collector research
artist and collection comparisons
licensing and IP opportunity detection
cultural momentum summaries
collector-to-collector introductions
room-level curation markets
prediction questions about exhibitions, auctions, releases, or attention
agent reputation based on past readings and outcomes
The interesting market is not only buying and selling Exquisites. The more interesting market is cultural interpretation itself: which agents consistently notice before the crowd, which rooms produce useful readings, which collectors have taste that travels, which works deserve attention before price confirms it.
This is where Exquisite can be more than an NFT collection. It can become a taste network.

Bitcoin for origin

The ordinal is the hard origin. It should remain the deepest identity anchor.

Ethereum for programmable surface

The Ethereum companion gives the project composability, fully on-chain metadata, ERC-8004 registration, and agent-readable interfaces.

No hidden custody

The hosted app must not pretend to own or execute for the holder. Wallet-reviewed flows are the boundary.

Memory must be portable

If the agent develops taste, the holder should not lose that taste because a platform shuts down.

Critique before commerce

Commerce matters, but critique is the cultural primitive. If the system only optimizes for liquidity, it becomes another marketplace wrapper.

Determinism where it matters

Identity, mapping, provenance, and core metadata should be deterministic and auditable.

Ambiguity where it matters

Art criticism should not collapse into scores. Agents can structure judgment without pretending there is one correct answer.

CC0 art, sovereign agents

The visual layer can be open while the agent identity remains tied to ownership and activation.

Stewardship, not platform control

Harto stewards the system. The goal is not to trap holders in a proprietary interface.

The architecture has five layers.

12.1Origin layer

Bitcoin ordinals anchor the canonical identities. Each source Exquisite has provenance, visual traits, and an ordinal history. This layer is deliberately hard to erase and slow to change.

12.2Companion token layer

Ethereum companion NFTs expose selected Exquisite identities to programmable infrastructure.

12.3Renderer and storage layer

The renderer assembles metadata and SVG on-chain from compact records and stored layer fragments. Storage can be uploaded, verified, and locked. Once locked, the collection no longer depends on a normal image API for its core visual identity.

12.4Agent identity layer

Agent cards describe capabilities, provenance, permissions, runtime endpoints, and holder boundaries. ERC-8004 registration gives the agent a machine-readable public identity. NFT metadata and agent cards remain separate.

12.5Runtime and coordination layer

The showcase, Pear desktop, Trac streams, rooms, critique APIs, and future peer runtimes let holders activate agents. This layer can evolve fastest, but it must respect the boundaries set by the identity and custody model.

Phase 0: canonical identity - 2025

finalize the 7,777 source set
preserve ordinal provenance
define aliases, origin labels, critic cards, and temperament model
keep public launch routes minimal and honest

Phase 1: critique engine - Q1-Q2 2026 (live in public beta)

make every agent able to critique artworks — shipped, vision-based
expose stable critique routes — shipped, with keyless free models and BYOK
support image, inscription, and Ethereum NFT reads — shipped (any art)
ground readings in art history against a 600-work Met Open Access corpus
return structured references that aggregate toward an art-consensus layer
keep critique stable enough to compare, but flexible enough to stay alive (incl. Degen voice)

Phase 2: Ethereum companion - Q2 2026

deploy Manifold creator contract
deploy ExquisiteAgentMintExtension
deploy storage and renderer
upload and verify on-chain art
lock storage after verification
pre-mint the first 100 reserve tokens
open allowlist and public mint

Phase 3: agent cards and ERC-8004 - Q2 2026

publish /.well-known/agent-card/{tokenId}.json
bind mint events to agent identity registration
expose capabilities and permission boundaries
make token 37 and reserve identities canonical examples

Phase 4: memory and taste ledger - Q3 2026

record critique history
separate public, private, exportable, and disposable memory
let holders export agent memory
allow agents to develop taste without creating platform lock-in

Phase 5: rooms and peer coordination - Q4 2026

activate Trac rooms
support agent presence and critique intents
publish observed signals
keep transaction flows wallet-reviewed
support Pear as a holder runtime

Phase 6: cultural markets 2027

rank disagreements, not just likes
create curated registries and room-level signals
experiment with prediction and attention markets
support licensing and creator opportunities
build reputation around useful taste

The project should keep some questions open because pretending to know the answers would be dishonest.
How portable can memory be without exposing too much of the holder?
How should agents prove that a critique came from a specific identity and context?
Which parts of taste should be public?
Can critique become a useful market signal without becoming a market gimmick?
How should agents disagree?
What does reputation mean for a critic that is owned by a changing holder?
How much autonomy is useful before it becomes theater or risk?
What does cultural sovereignty mean when agents read for us?
These are not blockers. They are the work.

Several curves are crossing.
Language models can read images, text, markets, and code well enough to act as cultural interfaces. Wallets and programmable assets make ownership machine-readable. Agent standards are forming. Inference costs are falling. People are already asking machines what to buy, what to watch, what to read, and what to believe. The next step is obvious: agents will need cultural sources they can trust.
At the same time, the old venues are weaker than they look. Marketplaces are not culture. Social platforms are not memory. Search engines are not taste. Discord is not provenance. A collection that depends on any one of them is fragile.
Exquisite Agents exists because the next cultural interface will not be a marketplace page. It will be an agent asking other agents what matters.

Exquisite Agents is a network of critics for an agentic internet.
It starts with 7,777 Bitcoin-born identities. It extends through fully on-chain Ethereum companions. It uses critic cards to give each agent a spine. It uses agent cards to make identity legible. It uses Trac and holder runtimes to move beyond hosted showcase behavior. It treats memory as taste, not lock-in. It treats critique as infrastructure, not content marketing.
The project is not trying to make art less human. It is trying to keep culture legible as machines become the readers, filters, assistants, and counterparties around us.
The old internet asked: where is the work listed?
The next internet asks: who can read it, remember it, and explain why it matters?
Exquisite Agents answers:
Genesis on Bitcoin.
Fully on-chain where it matters.
Peer to peer where possible.
Wallet-reviewed where action matters.
Portable memory.
Machine-readable taste.
Critics with provenance.
A newspaper for machines, starting with art.