A nightly council that learns, with a person in the loop
Your ads, checked every night.
What it learns, it keeps.
A council of eight AI agents and three people reads your data, challenges its own findings, and writes down what it learned as a short check that a person approves. Replay four recorded nights: the second night reads the first night's lessons and corrects its verdict. A stored lesson changes what a later night checks; it does not retrain a model.
- Calls to real tools
- Lessons a person approves
- No write without a human
Replay one recorded night of the council. It starts flat, the way a graph of who talked to whom looks. Then lift it into the room, to see where every answer came from.
Recorded run · client names masked · numbers rescaled · dates shiftedApprove and Return play the replay only; nothing is sent or stored.Space play · ← → 2 s · G gate
Meet the council
Eight AI agents. Three seats for people.
Each member owns a job, reads from its own systems, and shows its status honestly. The agents chair, measure, audit, research, write, argue, check and stage. The people review, fix and decide.
Recursive improvement, with a person in the loop
A night's lessons change the next night's checks.
After a run, the council writes what it learned as one short line: a check to run, its evidence and the night it came from. A person approves the text before anything is stored. The next night reads the stored lessons first. It changes what gets checked; it does not retrain a model.
- 1A findingAn agent reads real systems and says what it found, and what it could not see.AI agent
- 2Critic challenges itCritic argues how the finding fails: a lead is not a cause, a gap is not a no.AI agent
- 3A person approvesNothing is stored without a named person approving the text.Person
- 4A lesson is storedOne line: the check, its evidence, the night it came from.File
- 5Read the next nightAtlas reads the lessons before it frames the question.AI agent
Step 5 feeds step 1: the next night checks what the last one learned.
Four recorded nights
One connected-TV campaign, read-only, across four nights. Client names masked, numbers rescaled; dates in the four nights are shifted. The numbers below are counted from the recorded runs; 3 of the 28 calls read the lesson file, not an outside system. During the nights no write tool was called; after night four one approved write created a private page of lessons.
Lessons in memory
Skills from MCP
Skills you can check, tool by tool.
Every skill is bound to real tools on real MCP servers. The status is what the tool did when we probed it in early October 2026, not what we hope.
Designed to check, then decide
Four nights, step by step.
The nightly council console
Demo it. Replay it. Open its memory.
The replay at the top is one view of the product's own console. Open the console to walk someone through a night: every call on the replay bar, who did what and when, the council's channel, the 3D room and its 2D plan, every member's skills and tasks, the connections, the decision that waits for a person, and the Memory tab with every lesson and when it was read.
Space play · ← → 2 s · , . previous or next step · G the gate · in 3D, P flattens or lifts the room
The same recorded read-only nights, the same masked data. Approve and Return play the replay; nothing is sent or stored.
- LadderOne lane per member: each call, its server and its status, in time.
- ThreadThe council's channel, the proposal and the gate, as messages.
- 3DThe room with servers, the event log and the people, or its flat 2D plan.
- Skills, Tasks, Memory, ConnectionsWhat each member can do, is doing, what the council remembers, and what still waits for access.
The film
Four nights in one short film.
The same recorded nights, cut into a film: a campaign that looks wrong, a council that checks and disagrees, a person who approves, and the lesson that changes the next night.
Concept: film in production The film is being cut from these four nights. It is not published yet; the replay above is the same material.
720p · recorded runs, client names masked, numbers rescaled, dates shifted · voice and music: draft
Connected today
Where access allows.
Limited means some tools answer. Needs authorization means we're waiting for a consent. Planned means not built. Probed in early October 2026.
The human gate
Nothing ships without you.
Every recorded night ends with one proposal and the criteria to accept it. Maya reviews it; you approve or return it. Guard's rule: no write without a human. Enforcing that rule at a gateway is on the roadmap; in these runs no write tool was called at all.
- Read-only first: we connect only what you authorize.
- Anything that could move spend waits for a named approval.
- A lesson is stored only after a named person approves its text.
Read-only pilot · approvals stay with you
Design-partner pilot
Give us 30 nights.
We start read-only on your own data and show you a replay every morning. You decide what, if anything, changes.
- Week 1 · connect, read-only. We connect the systems you authorize: your MMP, ad platforms and spend records. Nothing writes.
- Nights 2 to 30 · a replay every morning. What the council read, what it found, what was blocked, and one proposal with acceptance criteria. Each night starts from the lessons a person approved before.
- You hold the gate. Anything that could move spend waits for your approval. Return it with a note and the council revises.
FeedOS is in private beta. We reply to every application by hand within a day.
FAQ
The honest answers.
Does the council run on its own?
Not yet. This page replays four recorded, read-only nights. The tool calls happened; the agents' messages were written from the recorded results. A council that runs every night on your data is what the pilot builds toward.
What does "recursive improvement" mean here?
After a night, the council writes lessons: each one a short check to run, with its evidence and the night it came from. A person approves the text before it is stored, and the next night reads the stored lessons first. So a lesson changes what a later job checks. It does not retrain a model, and nothing is stored without a named person. In the recorded nights, the second night reads the first night's lessons and downgrades its own pacing verdict after reading the client's words.
Where are the lessons stored?
In a plain file in the repository (one line per lesson) and in a private Notion page. Today only the owner can read them. Concept: a shared team knowledge base is planned, not built.
Is the data real?
The tool calls and tool names are real and the results come from real calls. Client, campaign and partner names are changed, the numbers are scaled and the dates in the four nights are shifted, so the figures are illustrative. "Northwind Cash" is a composite.
Who are the people in the portraits?
Nobody: they are AI-generated images of fictional people. The sparkle badge marks AI agents, the person badge marks the three human seats (a FeedOS strategist, ad ops, and you, the approver).
Will it write to my ad accounts?
Not without you. Guard's rule is no write without a human, and every write skill is marked "Write · needs approval". The Beeswax connector we used exposes write-capable methods, and nothing enforces read-only at a gateway yet: it holds only because no agent called them. Enforcing it is planned. In the recorded nights no write tool was called.
What is connected today?
See Connected today: FeedMob's own data and database, Mobius, Notion, Fireflies and GitHub work; Beeswax, AppsFlyer and TikTok are limited; Singular and Meta wait for consent; OpenAI Ads is planned. We show it as it is.
What does a pilot look like?
Thirty nights, read-only first, on the systems you authorize. Every morning you get the replay and at most one proposal with acceptance criteria. Anything that could move spend waits for you.
Why 3D?
Because the question that matters is not only who said what, but where it came from. In the room every server, database and API is an object with a status light, and you can watch each call leave an agent, reach a system and come back, or not.