How it works
You ask, your agents do the work, you approve it. WilsonOS is your workspace and five things under it. The workspace is the one place you and your agents both make things, and each of you can see what the other made. The five pillars are what make it work. Every piece of work — a book launch, a week of Pinterest posts, a grant application, a personal goal you keep slipping on — then moves through the same five steps. Learn the pillars and the steps once, and the rest of these docs are detail on each.
Your workspace
Section titled “Your workspace”Where the work, and the state of the work, both live. It is organised by what you are doing rather than by screen: Talk to any agent, See goals and what is waiting on you, Review what an agent made next to the task it came from, Make things yourself in the Studio, Run an engine from its cockpit, and Keep your catalogues of contacts, products, files and skills. Each is described in Your workspace.
The five pillars
Section titled “The five pillars”01 · Your agents — who. One agent per brand, set up your way, able to call whichever helper the work needs. Start as many conversations with it as you like; each runs in its own isolated workspace with a short-lived pass, never a provider key. See Your agents.
02 · Goals, tasks, artifacts — what. A goal is the why: a number you are moving, a practice you are keeping, or a state you are becoming. A task is a few hours of work that serves one. Everything an agent makes lands as a version on the task it came from. See Tasks and the Workbench.
03 · Capabilities — how. Skills are know-how; workflows chain skills together; tools and connections are what an agent can reach. Trust belongs to the skill or workflow, not the agent, and an upgrade ships only after Skill Showdown proves it better. See Skills.
04 · Your shared brain — with what. Your brand, customers, products and pages, plus every decision, fact and event that affects the brand. Anyone plugged into the brand, person or agent, reads it and can add to it, and every entry shows who added it. See What your agent remembers.
05 · Engines — what runs without you. The repeating parts of the business, like your ads and your pins. Each cycle an engine calls what will work, does it, then checks what happened against what it predicted. That gap is what teaches it. See Channel engines.
Under all five sit the foundations: who can see and do what, and one writer for every kind of record. See Foundations.
The five steps
Section titled “The five steps”The pillars are what WilsonOS is made of. The five steps are how a piece of work moves through them: the agent’s loop, in plain words.
-
Tell it your world
Section titled “Tell it your world”Who you are and how you like to work, your goals and commitments, your businesses with their brands, products, offers and customers. Described once and kept current. The stable parts load at the start of every conversation; the detail is pulled in when a task needs it. Your personal area is coordinated by Mia and stays private to you.
You decide: what is in your world, and who may see what.
-
Talk to your team
Section titled “Talk to your team”State an intention — “help more parents discover this book”, “get the grant application in by Friday” — and say what success looks like. The agent already knows the context, asks only what it needs, and when something is genuinely your call it stops and asks with concrete options. Recurring work is set up once as a cadence.
You decide: the intention, the definition of success, the answers.
-
It does the work
Section titled “It does the work”The agent breaks the work into tasks with clear outputs and handoffs, picks the skill or workflow that fits and says which, loads what each task needs before it needs it, and runs the routine checks. Progress streams to you as it happens. Anything that costs money stops and asks, including inside a scheduled run.
You decide: whether to spend, and any mid-course correction.
-
You decide
Section titled “You decide”Results land as versions with their checks already run and shown. You see evidence, exceptions, and the decisions that are yours: approve a version, adjust it with focused controls, open a Studio when the work deserves it, publish through a connection you authorised. Every “yes” is on a specific thing.
You decide: what is good enough, and what goes out.
-
It learns, you measure
Section titled “It learns, you measure”Decisions carry forward. Ratings attach to the version they were given on. Business results arrive with their source, freshness and limits visible. A capability improves by proposal, test, review and promotion, never by reacting to one good result, and a weekly review brings you the proposals worth accepting.
You decide: which improvements to accept.
Why it works
Section titled “Why it works”One principle explains the shape of the whole system:
Give the agent the right information, know-how, and tools for the task at hand, at the time it needs them.
| What the agent gets | Which part of the system supplies it |
|---|---|
| A clear intention and a definition of success | Your request, or the goal it serves |
| Work broken into tasks with outputs and handoffs | Workflows |
| The right information, loaded just before it is needed | The shared brain and session start |
| The know-how for this exact job | Skills |
| Tools and permissions scoped to the task | Connections |
| Checks against the intended outcome | Review in the Workbench |
| Feedback that improves the next run | The improvement contract |
None of these is useful alone. A skill without context produces generic work; context without a method produces improvisation; both without checks produce output you have to re-read; and all of it without feedback never gets better. Together they protect you from a session overloaded at the start, a long task that drifts, a failure hidden inside a finished-looking result, and “improvement” driven by one anecdote.
A day with WilsonOS
Section titled “A day with WilsonOS”Sam runs a small publishing business with two brands and a life that does not fit in a spreadsheet. This is a Tuesday, and nothing in it is heroic. Items marked (intended) are designed but not yet available.
7:40 — priorities, not a to-do list. Mia’s morning note is already on Sam’s phone: three things that matter today, a commitment due from last week, and two business items that need a decision. Everything else is moving without Sam. Sam replies, “Push the podcast prep to Thursday, I want the morning for the activity-book cover,” and Mia records the decision for Thursday’s plan.
8:15 — a decision with evidence. Overnight, Scout finished three cover options for Code Pineapple’s holiday activity book, on a quote Sam approved in chat the evening before. They arrive with their checks already run:
Three options, generated on the quote you approved yesterday. Checks run on each:
- Brand paletteA B C
- Title legible at thumbnail sizeA B C
- Audience fit — leans on two phrases your customers useA B C
Sam compares A and C, comments “warmer sky” on C’s version, and approves A. The comment stays on that version, so Scout’s next pass addresses it there.
9:00 — a small adjustment. The cover needs a square version for the store listing. Its card offers a mini app with three controls: crop focus, background extension and a title-safe margin. Sam adjusts the crop and saves, and it becomes a new version on the same task.
10:30 — business work moves anyway. Prime is running RiseBlox’s weekly Pinterest cadence: drafting pins from the customers’ own language, checking each against brand voice and the platform’s rules, and queuing them. It stops once, because six images cost money. Sam reads the quote and says “go”.
14:00 — a tracker, not a spreadsheet. Mia drafted a conference proposal from the brand’s positioning and Sam’s notes. Sam edits two paragraphs and marks it ready, then submits it through the conference’s own form and records that on the opportunity, which moves to submitted with a follow-up scheduled (intended: the application tracker). Ready and submitted are separate states because only Sam can do the second.
16:30 — results, honestly. Prime’s afternoon note shows yesterday’s pins by impressions and saves, and says plainly that clicks cannot yet be linked to store sales for this title, so it will not claim any. One pin family is clearly ahead; Prime proposes three more in that direction. Sam says yes to the pins and no to changing the pinning skill. That is how the skill stays good.
17:15 — the day closes itself. Mia’s evening review lists what shipped, what waits on Sam tomorrow, and the commitment Sam kept. Sam adds a preference in passing, “always show me cover options on a dark background too”, and it is recorded where Scout will find it next time. Sam spends the evening on a novel, for no business reason at all.
What the day shows. Sam delegated the routine and kept the judgement. Sam saw evidence, exceptions and decisions, not the whole board. Every “yes” was on a specific thing, and spend stopped and asked even inside a scheduled cadence. Personal and business shared one system without sharing visibility.
What stays constant
Section titled “What stays constant”- You never choose a tool. You state an intention and what success looks like; the agent chooses the method and says which one.
- Every “yes” is on a specific thing: a version, a price, a plan, a spend.
- Guardrails do not depend on you watching. Cost approval, brand rules and checks are the same for a chat request and a scheduled run.
- Success is capacity, not volume. The measure is better results and a less overwhelmed you.
- The loop closes. Results feed the next intention. That is what makes it a system rather than a very good assistant.
For builders — the same loop as eight stations
Engineers describe the five steps as eight stations, each answering one question and each tied to the files and tables that do it. How work gets done walks five real requests through them.
| # | Station | The question it answers | Pillar |
|---|---|---|---|
| 1 | Trigger | What starts the work: a message, a schedule, or an event? | 01 Agents (a message, a schedule), 05 Engines (a signal) |
| 2 | Workflow | Which workflow does this activity run? There is always one. | 03 Capabilities |
| 3 | Context | What does the agent already know when it wakes? | 04 Shared brain |
| 4 | Skill | Which know-how loads, just before the step that needs it? | 03 Capabilities |
| 5 | Brand | Whose voice, look and customers does the output follow? | 04 Shared brain |
| 6 | Action | Make something, pull a lever, or ask a person? | 01 Agents, 02 for what gets made, 05 Engines for levers |
| 7 | Track | Where does the result live, and who decides it is good? | 02 Goals, tasks, artifacts |
| 8 | Learn | What changes next time, from the rating and from the outcome? | 03 Capabilities (from your rating), 05 Engines (from the outcome) |
Learn takes two inputs: a rating, which says whether the work was made well, and an outcome, which says whether it worked. What it produces (a better skill, an updated playbook, a sharper forecast) is what the next request already knows, which is why this is a loop and not a pipeline.
An engine runs the same loop, smaller and faster: target, predict, act, measure, learn, and how much rope it has earned. Use the stations when you need to know when something happens to a piece of work, and the pillars when you need to know where it lives.
Related
Sources: Don's product direction, 2026-09-16 (task 410 review), input/WilsonOS_Overview.html §02, §04, §05, docs/architecture/overview.md, docs/architecture/concept-map.json, docs/DECISIONS.md (ADR-018, ADR-028, ADR-029), docs/architecture/skills-and-flows.md (load-once rule, flow authoring standard), docs/architecture/memory.md (boot tiers), docs/architecture/engine.md, docs/wilsonos-app/PRD.md §3.5 (Mia's cadences), WilsonOS Team Brief (prototype), "The shape" and "The five underneath" — absorbed (task 494) · Last verified 2026-10-06

