For Meta · the roadmap · August 2026
Ads are the business. Signal is the constraint.
The roadmap we would build with Meta, published in the open. Ads are effectively all of Meta's revenue, ad growth now comes from price per ad, price per ad is relevance, and relevance is signal. Signal is the input getting harder to source. Everything below follows from that.
Meta today
Start with the P&L, not our product.
Q2 2026, from Meta's own disclosure.
Up 28% year over year.
Up 27%. Effectively the whole company.
The growth engine. Relevance, priced.
Inventory growth, which cannot compound forever.
Quarterly operating loss, on $431M of revenue.
$31.1B in Q2 alone.
What the numbers say
Growth is coming from relevance, and relevance is signal.
Ad revenue is $59.4B of $60.8B. The glasses, the models and the capex are all funded by that line. So the real question for anyone building with Meta is what moves it.
Impressions grew 14% and price per ad grew 12%. Impressions are inventory, and inventory has a ceiling. Price per ad is relevance, and relevance runs on signal. That is the half of the growth that can keep compounding, and its input is getting harder to source every year.
There are two answers to a signal shortage. Infer harder, which regulators, operating systems and people are all pushing back on. Or be given the signal. A declared intent from someone who chose to share it is more accurate than any inference, and it gets easier to obtain as privacy law tightens.
That is the layer we build. Meta has the demand, the auction, the ranking and the reach. What no one can manufacture is a person's willingness to say what they want. That has to be earned by someone standing on the person's side of the line.
Where Meta is going
What Meta says it is building.
The public record as of August 2026, from Meta's earnings call and announcements.
Personal superintelligence, for billions
An AI that knows a person's own history, interests, relationships and goals. The company's north star, with 2026 called a big year for delivering it.
Agents that run continuously
Within five years, billions of people relying on autonomous agents working in the background on their behalf.
The four things those agents will handle
Personal finances. Health. Household logistics. Relationships. In order, the four most sensitive data sets a person has.
WhatsApp as the primary interface
The agent lives where the conversations are. Family of Apps Other revenue, largely business messaging, grew 73% this quarter.
An ads stack that is AI end to end
Andromeda retrieves candidates, Lattice ranks, and GEM sits above both, trained at LLM scale and predicting the sequence of actions around an ad.
Glasses as the surface
A target of ten million pairs of AI glasses in 2026, a display line arriving through the autumn, and Muse Spark described as running on device.
Where we meet
One thing capex cannot buy.
Read those six lines together and one requirement runs under all of them. An agent that manages someone's finances, health, household and relationships needs that person to hand over the four most sensitive data sets they have. An ads model that predicts the sequence around a purchase gets far better with declared context. Glasses that see what the wearer sees need permission from the wearer and goodwill from everyone else in the room.
Compute can be bought, models trained, distribution already exists. A person's willingness to share what is going on in their life has to be earned.
Meta's personal agent runs on Meta's infrastructure. Ours runs on the person's. Those are different products, and they serve different people: many Americans will never hand that data to a platform at any level of capability. Serving them takes someone whose whole business is being on their side. That is the segment we open, and together we reach everyone.
You are building the agent. We are building the reason a cautious person lets one into their life.
The roadmap
Five workstreams. One starts this week.
Each carries what we would ask for and the number that would tell us it worked.
Design partner on Muse Spark 1.1
Ready nowThis quarter. Startable this week.
The Meta Model API is OpenAI-compatible, and bring-your-own-API is already first-class in 🤫 Private Agent One: the person's key, their quota, their hardware. So an American can point their own agent at Muse Spark 1.1 today. We would love to do it properly, as a design partner, with the feedback loop pointed back at Meta.
Design-partner access and a channel to the model team.
Agent One shipping Muse Spark as a selectable provider, with real usage from real people on their own keys.
Consented signal for Ads
NextOne surface, one quarter.
A person publishes a declared intent, not an inferred profile. Advertisers bid for permission to reach that intent. Matching happens on the person's device and no identity is transferred. Signal given rather than taken, and it gets better as privacy law tightens.
One surface, one market, and an agreed measure of lift against today's baseline.
Price per ad on the pilot cohort against control. If it does not move, we both learn that quickly.
Agent One on Ray-Ban Display
NextAfter 00, with a hardware conversation.
Glasses see what the person sees, so trust matters most here. An agent that holds context locally and shares per field, with a receipt, is the difference between a device someone wears among other people and one they leave at home.
Developer access to the glasses platform.
A working on-device agent loop with consent and receipts, on hardware.
Business messaging that respects the person
NextFollows 01.
Meta has named WhatsApp as the primary interface for personal agents. Every business conversation is a business asking a person for something. The consent handshake, the scope and the receipt belong there natively, and a business that can prove it only received what it was granted is a business people answer.
A pilot with a small set of business-messaging customers.
Reply rate and complaint rate against today's baseline.
Compute at the edge
NextLong horizon.
A distributed grid of owner-operated supercomputers is a different cost curve for a specific class of workload: inference close to the person, on hardware somebody already paid for. It complements a data center; it does not replace one.
A technical conversation about which workloads fit.
Cost per inference at the edge against the data center, measured plainly, including where we lose.
Workstream 00
Muse Spark 1.1, with the person holding the key.
Muse Spark 1.1 plans, calls tools, drives interfaces and delegates to subagents. That is an agent acting on someone's behalf, and it needs their real context: their calendar, their money, their health, their family. Where that context lives, what is kept and what can be taken back is the question that decides whether anyone hands it over.
🤫 Private Agent One answers it, at no cost to Meta to try. Bring your own API and bring your own compute are native: the person supplies the key, the quota is theirs, and the agent runs on hardware they own. The model is called, not fed.
Meta gets
Usage, a feedback loop from an agent doing real work for real people, and proof that the model can be trusted with a life.
The person gets
Frontier capability on their own terms. Their key, their hardware, their receipts, and the freedom to switch models any day.
We get
To be judged on whether people actually trust it, the only test of a consent layer that matters.
Boundaries
What we are not asking for.
- No change to Meta's business model. Ads pay for everything Meta does, and that is fine.
- No exclusivity, in either direction.
- No requirement to adopt our protocol. If a better consent standard exists, we would rather use it.
- No announcement. We would rather ship something small that works than sign something large that does not.
Where it stands today: workstream 00 needs only an API already in public preview, which is why it is ready now. The consented signal layer the others depend on is partly built. The consent protocol, per-field grants and receipts are real; the auction, the advertisers and the payouts are what we would build together.