Consumption Agents: The Missing Half of the AI Agent Landscape

First, what this post is not about: everyone is talking about agentic commerce — AI that shops for you, pays for you, negotiates for you. This post is about the half that got skipped. It represents your attention, not your wallet.

Over the past year I turned myself into a one-person product studio run with a fleet of agents. They write code, do research, produce reports — the output is real. And that’s exactly how I ran into a problem nobody seems to discuss: the more they produce, the more miserably I consume. Ten research reports sit in a folder waiting for me. Three agents’ deliverables wait for review. Every feed I subscribe to is now “AI-enhanced” and posting faster. My agent fleet solved production, then handed the entire consumption bill to my biological brain.

I stared at that discomfort long enough to realize it isn’t my personal problem. It’s a structural hole in the agent landscape.

A dichotomy that had no name

Every widely deployed AI agent today — coding agents, data agents, research assistants — is a production agent: intent goes in, an artifact comes out, and the world holds one more thing. Writing code, drafting reports, generating articles: all of it is “making the thing.”

The missing half I call the consumption agent: it adds no new information to the world. It takes what already exists — articles, feeds, reports, and increasingly the output of production agents — and performs selection, transformation, and orchestration so it fits this person, in this moment: your goal, your time, your device, your cognitive state. It digests things for you, or blocks them for you.

Production agents change the state of the world. Consumption agents change your interface to it.

Before writing this I ran four independent literature searches: nobody classifies agents along this production/consumption axis. The phrase “consumption agent” only appears in BI vendor docs (querying semantic data models) and in energy contexts — nothing to do with attention. The naming vacancy is real. It’s also closing.

The criterion: don’t ask what it does, test the output

“Summarizers already exist,” you might say. “Chatbots already help you consume.” So the criterion has to be nailed down. Consumption agents generate tokens constantly — the dividing line can’t be “does it generate.” It’s three tests on the output:

Audience test — made for the public record (production), or for one person in one moment (consumption)?
Lifetime test — must it stay correct and citable (production), or is it disposable by design, legitimately different tomorrow (consumption)?
Information test — does the world gain a new fact (production), or does an existing fact gain a better fit to one mind (consumption)?

Run some calibration cases through it and you get deliciously counterintuitive results:

The spam filter is history’s first mass-deployed consumption agent — it says “no” to content on your behalf, serves only you, adds nothing to the world. It just isn’t very smart.
TikTok’s For You feed is a consumption agent too — owned by the platform, not by you. It senses your state, but its loyalty runs to a watch-time metric, and it will never advise you to stop scrolling.
NotebookLM is close, but does transformation only (long text in, podcast out) — no selection, no “when and at what pace” orchestration.
ChatGPT in its default use is production — you ask it to write; the world gets one more artifact.

Content is source code. You are the target platform.

The most explanatory metaphor here is compilation.

For eighty years, content has been AOT — compiled ahead of time. The author locks in one version; a million people consume that same version; every adaptation cost is paid in reader attention — skimming, abandoning, doomscrolling, grinding through.

A consumption agent makes content JIT: meaning is produced once; the experience is compiled at the moment of consumption. The same deep-dive report becomes a 15-minute podcast during your commute, a one-screen brief between meetings, a conversation you can interrupt on the weekend. Used once, discarded, recompiled differently tomorrow.

Content is source code, the person is the target platform, and the consumption agent is a just-in-time compiler that answers only to you.

Note the last clause — “answers only to you.” That’s the life-or-death question of this whole direction, and we’ll get to it. First: why is this possible now?

An eighty-year-old dream, two bricks short

The idea is not new; the window is. Bush’s Memex (1945) was consumption assistance from start to finish. Maes’s interface agents (1994) put the motivation right in the title: information overload. Negroponte’s Daily Me (1995) described “your interface agent reading every newswire and constructing a personalized summary” in nearly these exact words. Five generations tried; each hit a different wall — no AI that understood semantics, no way to generate variants, no business model.

Then LLMs dissolved the two hardest bricks at once: understanding and generation. In November 2025, Google shipped generative interfaces with Gemini 3 — custom UI generated per prompt. Within a single year, three separate presentation-layer protocols appeared: “how agents render to humans” got institutionalized as its own problem. The empirical base arrived too: generated interfaces beat Markdown in user preference 82.8% of the time; re-rendering learning material lifted retention by 11 points; designers agree with each other on “what UI is good” at only κ=0.25 — one-size-fits-all fails on the data.

Every component is mature. Nobody has assembled the whole: selection (L1) + transformation (L2) + orchestration (L3), resident, loyal to the user — that complete stack exists nowhere: not commercial, not open-source, not academic. Recommenders do L1 only. NotebookLM does L2 only. L3 — when to give it to you, at what rhythm, when to interrupt, when to stay silent — exists only in fragments. What’s missing isn’t capability. It’s the scheduling intelligence that knows what this moment should compile to.

That’s the gap. That’s also the opportunity.

The life-or-death line isn’t technical. It’s loyalty.

If I stopped here, this would be an optimistic tech essay. But there’s a question this direction cannot dodge: whom does the rendering layer work for?

We already ran this experiment once, for twenty years. It’s called the feed. When the rendering layer works for the platform and optimizes engagement, you get addiction design and echo chambers — not because engineers are evil, but because whoever pays the agent decides what it optimizes. I sort the possible forms into three archetypes:

Archetype	Behavior	Optimizes	Business model
Dealer	amplifies what you crave right now	engagement / time-on-app	advertising
Butler	obeys your explicit instructions	task completion	subscription / one-time
Coach	serves your long-term intent, dares to defy your comfort	long-term goals	user-funded + on-device + deep trust

Ad-funded rendering degrades into the dealer — that’s not a moral judgment, it’s incentive structure. So the conclusion is hard: user-funded + on-device + auditable is the only soil where a user-loyal consumption agent can grow. Which happens to be exactly what ad-dependent giants structurally cannot do (agent-mediated consumption kills advertising — it’s self-revolution), and therefore the one door left open for new players.

There’s a second, sneakier channel to guard: the training signal. Even at 100% subscription revenue, a system that learns only from your taps, skips, and completions is running the engagement gradient — dealer dynamics don’t need ads; they grow out of the data loop. So the user model must split into two layers: a descriptive layer learned from behavior, and a normative layer — your written declarations, never overwritten by behavioral data. If you said “less short-form video,” ten slipped taps must not repeal it.

From these tensions I set three non-negotiable axioms:

Loyalty must be auditable — proven by architecture, not promised; audited on both channels: where the money flows, and what the model learns from.
Every rendering must anchor back — one click back to the canonical source. Rendering is glass, not a wall; that’s also basic ethics toward upstream creators.
Know when not to transform — recognize form-is-meaning content (poetry, legal text, material that needs desirable difficulty) and refuse to flatten it. It’s a dietitian, not a juicer.

What I’m building: OctoAgent

A concept that stays in an essay is just a manifesto. I’ve started building it as OctoAgent — named for the octopus architecture: one central brain + pluggable arms.

Central brain: a taste passport (a consumption-preference model you own and can carry), an intent compiler (“I’m exhausted tonight but want to keep up with AI” → rendering parameters), and the loyalty kernel;
Arms: concrete consumption scenes — an agent-deliverable inbox, a morning brief, a reading coach, a report translator…

A real octopus keeps about two-thirds of its neurons in its arms. So the arms come first: the first one I’m building for myself is the inbox that digests my agents’ daily output — the exact discomfort this post opened with. The central brain gets extracted from real use, not designed as a platform on day one.

The full concept — the eighty-year lineage, the product casualty list, the falsification ledger from the research — is an open seven-part English handbook (CC BY 4.0): github.com/Octo-o-o-o/consumption-agent. Every number traces to a primary source through the verification ledger, and the claims that died during research are preserved there too. The repo practices what it preaches.

Closing

Recommendation algorithms decide what you see. A consumption agent decides how you see it — and it answers only to you.

The production-side arms race will continue and the flood will rise. But on the interface side, for the first time in eighty years, the term “User-Agent” has a chance to mean what it says.

The category has a name now. Next, it needs a body.

Numbers in this post use the handbook’s evidence-graded figures (✅ adversarially verified / ◎ corrected after audit); full provenance in the verification ledger. This is the first post in the AI Watch series, and the starting point of the OctoAgent project.