Not on Demand. On Schedule.
In April, the team at Meta Analytics published a case study.
Not about Llama. Not about an external product. About an internal tool they built for themselves.
They called it a “second brain.” Sixty thousand people across Meta are using it — engineers, PMs, designers, legal, finance, communications, sales. The writeup described what the system actually does: tracks projects, reads meeting notes, surfaces connections, builds on prior conversations.
There’s one detail in the writeup that most coverage missed.
“Proactive agents run on schedules rather than waiting for prompts.”
Morning briefings. Automated meeting note processing. End-of-day digests. The system generates output before you open it. Before you formulate a question. Before you decide to ask.
That’s not a detail. That’s the design.
Every AI tool you have is waiting
Open ChatGPT. Nothing happens.
Open Notion AI. Nothing happens.
Open Copilot, Perplexity, Claude, any of them. Nothing happens until you type something.
This is the universal design pattern across the entire current generation of AI tools: reactive. You bring the context. You initiate the request. You phrase the question. The system responds. You assess the output. You decide what to do next. You track whether it was done.
At every step, the cognitive load lives with you.
This isn’t a model quality problem. The models are good. The reasoning is coherent. The outputs are often useful.
The problem is architectural. The trigger is always you.
What happens when the trigger is the system
Meta’s second brain doesn’t wait to be invoked.
At a defined time each morning, it processes what’s relevant and delivers a briefing. When a meeting ends, it processes the notes — automatically. When a project status changes, it surfaces the shift without being asked.
The user’s job is not to remember to ask. The system has already asked itself.
This changes the cognitive model entirely.
The reason the BCG brain fry study found 14% of AI users cognitively overloaded — and 26% in marketing — is not that those workers are doing more. It’s that supervision is ambient and open-ended. When did I last check the output? What’s been decided and what’s still open? What needs my attention today?
Those questions don’t have natural triggers. They don’t occur to you on a schedule. They surface at 11 PM when something slips through, or at 9 AM when someone asks why a loop didn’t close.
Reactive AI doesn’t solve this. It waits for you to remember to ask.
Proactive infrastructure solves this. It runs without you remembering anything.
The “Third Brain” and why it matters
The Meta writeup introduced another concept.
Beyond the individual second brain, they built a “Third Brain” — a team-level shared context layer. Individual workspaces feed into a shared knowledge layer. The team’s coordination overhead — status updates, handoff clarity, who owns what — moves into the system and out of the channel.
This is the coordination pattern at its logical next step.
The individual second brain reduces the cognitive load on one person. The team third brain reduces the coordination overhead of the group. The system tracks commitments across people, not just across tasks.
When someone says “I’ll follow up on this,” the loop doesn’t live in someone’s memory waiting to be asked about. It lives in the layer.
When a decision gets made in a meeting, it doesn’t evaporate into notes nobody reads. It gets absorbed and becomes part of the context for what happens next.
This is what coordination infrastructure actually looks like when it’s functioning.
Why your tools don’t do this
The current market for AI tools was built on the question-and-answer model.
It’s intuitive. It maps onto familiar behaviors — search, messaging, lookup. You have a question, you ask, you get an answer. The value is in the quality of the answer.
That model is useful. It also has a hard ceiling.
Because the coordination problem isn’t a question-answering problem. The open loops in your working memory aren’t questions you’re waiting to ask. They’re commitments with unknown states, decisions with unclear ownership, projects with unclosed actions. They don’t have a natural trigger.
You don’t open ChatGPT and ask: “What are the seventeen things I delegated this week that haven’t resolved yet?” The question requires context you’d have to supply. The answer requires tracking you haven’t done. The burden of formulating the question is itself a symptom of the problem.
Proactive infrastructure doesn’t need you to formulate the question. It already knows what it’s tracking. It surfaces the answer before you thought to look.
The design shift
This is not an incremental improvement to existing tools.
It’s a different architecture.
Reactive AI tools: you initiate, they respond. The context lives in your head until you extract it.
Proactive coordination infrastructure: the system initiates, it surfaces, it closes. The context lives in the layer.
The shift is from AI-as-tool to AI-as-operator. Not waiting to be invoked. Running the protocols it was built to run.
Meta had to build this internally because it doesn’t exist as a product. Sixty thousand employees. Some of the most AI-capable engineers in the world. Still had to build it themselves.
The question isn’t whether this architecture is better. Meta’s scale deployment answers that.
The question is: why isn’t it available as infrastructure for everyone else?
The gap
Every week, knowledge workers are adding more reactive AI tools. The BCG data says above four tools, cognitive strain reverses productivity gains. Not because the tools are bad. Because each tool requires you to know when to invoke it, what to ask it, and how to track what it produced.
The coordination overhead of managing reactive tools is itself a cognitive load problem.
The solution Meta built — proactive, scheduled, team-level, context-aware — is the pattern. They validated it at 60,000-person scale in April 2026.
The category doesn’t have a name yet. The product doesn’t exist as infrastructure yet.
But the design is clear:
Not on demand. On schedule.
The system runs the protocols. You handle what requires judgment.
Everything else should already be handled.
