Skip to content
Describes how WilsonOS is designed to work.What works today →

Agents

Pillar 01 · Agents

A COO agent for every business, and a chief of staff across all of them: colleagues with a name, a lane, and a defined set of things they know the moment they start.

A team of agents, each with a name, a role and a lane. Every business you run has its own COO agent, who owns that business’s work from research to publishing. Mia, your chief of staff, works across all of them and your personal life. Each agent starts every conversation already knowing you, its business and what is on its plate, and stops to ask before anything that is yours to decide.

Your agents3 shown · pick one
NameTitleBusiness
MiaChief of staffEverything, and your personal life
ScoutCOOCode Pineapple
PrimeCOORiseBlox
Add a business

A new business comes with its COO agent

  1. Describe the business: what it sells, who buys and why, how it sounds and looks.
  2. Its agent is created with a name, a personality and a lane: that one business, end to end.
  3. You review its profile and switch on the routines it proposes. Nothing runs until you do.

Mia

Chief of staff · works across every business

coachsystem stewardpersonal area
Owns
Your priorities, the weekly review, and keeping the plate honest
Personality
Warm and direct
Knows at the start
You, your goals, every agent's plate, what slipped
Routines
Morning priorities, evening review, the weekly review, and the overnight upkeep of every agent's memory
Asks before
Spending, publishing, or changing a rule

Her full profile is further down this page.

A labelled illustration of the idea, not the app's screen. Mia, Scout and Prime are real agents; other businesses' agents are left out of these docs. Setting up a new business and its agent is done once by the workspace owner today; creating both from one step is Planned.
What you do
  • Talk to the agent for a brand, and to Mia about everything across them.
  • Read an agent's profile before you rely on it: role, routines, connected services.
  • Approve the routines an agent proposes for itself.
  • Correct what it keeps — a fact, a decision, a preference.
What your agent does
  • Starts every conversation already knowing you, its brand and its plate.
  • Pulls in deeper context only when a task needs it.
  • Runs its scheduled routines and reports to the conversation you chose.
  • Stops and asks before spending, publishing or changing a brand rule.

The lanes do not overlap: one business, one agent, one voice. Business agents and Mia deliver business work and may improve the skills and workflows they use; an engineering agent builds the app itself and never writes brand content. An agent’s identity — who it is, how it starts, how it talks, which routines it owns — changes only with your approval. Renaming an agent keeps its history; deleting and recreating one starts a new identity that inherits nothing.

“Knows you, your brands, and what’s on your plate today” has to be true at the first message, on every surface. So every agent starts with the same lean bundle, and nothing else until a task asks for it.

Session bundle · one agent, one person
Fixed documents
identityWho the agent is and how it behaves
house rulesThe workspace's working discipline
your profileYou, your preferences, your hard rules
brand essenceThe brand's one-page summary
Live · assembled for you and this agent
open plateTasks assigned to this agent
feedbackReviews waiting on it
top factsA ranked digest of what matters most
recent eventsWhat happened lately, noise filtered
failed runsIts routines that failed in the last 48 h
last stretchA short story of what it was last doing
Both halves load before the first message. The bundle is small on purpose: it is paid on every conversation.

Everything else waits until it is needed: a brand’s customer language when a skill declares it, deeper brand documents when the agent follows a link from the essence, the results ledger for a weekly review. Each piece loads once per task, and nothing is preloaded “in case”. At the end of a task the agent records what should last — decisions, durable facts, a status — and those records are what the next conversation’s bundle is built from.

An agent is easiest to trust when you can see what it knows. Here is Mia’s.

M

Mia

coachsystem stewardno brand — cross-cutting
Role
Accountability coach for the owner; keeps the plate honest; runs the weekly review; proposes improvements to skills, workflows and the system.
Supports
The owner’s personal area and the system as a whole. Brand agents own their brands; Mia can say what each is working on without loading their context.
Routines
20 enabled, 4 paused (her registry, 16 Sep 2026): morning priorities and evening review, the weekly review, nightly memory consolidation for every agent, drift checks, metric snapshots, a weekly feedback digest, and a monthly skill-catalogue review. Each names its brand scope, its destination, and the credentials it may use.
Current work
Illustration: three open items, one feedback flag from last week’s review, no failed routines in two days.
Mia’s profile — a labelled illustration, not a screenshot. The routine count is real on the date shown; the current-work line is illustrative.

Retrieves when a task needs it

  • A brand’s documents, when that brand comes up
  • The results ledger, for a weekly review
  • Your recent turns, when you ask to resume
  • A skill’s reference material, when she runs it

Can act on

  • Any skill or workflow, saying which
  • Tasks, through the task writers
  • Memory, scoped to you and her
  • A routine for herself — you enable it
  • Spend — only after you say go

Keeps afterwards

  • Facts and decisions from the session, scoped to you
  • A refreshed summary and fact digest, overnight
  • Nothing in shared brand knowledge unless you approve it in the weekly review

What an agent knows depends on who is present, because private context belongs to a person, not to the agent.

Private conversation

  • Identity, house rules, her plate
  • Your profile and hard rules
  • Your facts and summary, scoped to you and Mia
  • Acts with your rights, on your brands

Shared room

  • Identity, house rules, her plate
  • Anyone’s profile or private memory
  • The room’s own conversation memory
  • Acts with the sender’s rights each turn, on the room’s brand only

Scheduled run

  • Identity, house rules, her plate
  • Anyone’s private memory, and it cannot write any
  • The brand documents the routine declares
  • Acts on the declared brand, with only the named credentials

Chat is the trigger you see. Much of an agent team’s value is in work that starts without you: the morning brief, the weekly review, the campaign that posts on schedule, the nightly performance pull. Work starts three ways — you ask, a schedule fires, or a connected channel reports something — and whichever it was, the same things hold: a run record is written before anything executes, the work happens under a credential scoped to one brand and one run, spending waits for approval, and the outcome is written down.

Schedules. Every agent owns a registry of its routines: what fires, when, in which brand, and where the result goes. Each entry declares its brand and any outside access it needs rather than inheriting broad access, and names the conversation it reports to; without one it fails loudly instead of posting somewhere silently. Agents may add routines for themselves, switched off until you enable them, and pausing is preferred to deleting.

Signals. Some work starts because the world moved: new performance data landed, a post went live, a listing changed. A channel engine pulls those signals in, works out what changed, and proposes actions, which arrive in a conversation like anything else. See Engines.

The same guardrails, unwatched. An unattended run is never less careful than a conversation. It names the skill it ran, quotes before spending, applies the brand’s rules, and writes memory through the same scoped writers. A watchdog compares each routine’s expected cadence with what actually fired, and failures from the last two days appear in the agent’s own start-up bundle, so it knows and can tell you. Silence is treated as a defect, not as calm.

For builders — how an agent actually runs

One always-on process per agent, one runtime per conversation. Each agent is a bridge the app talks to. A message is routed to a runtime for that conversation: its own model session, working area, system account and short-lived credential. Two conversations with the same agent are two runtimes, so stopping one leaves the other running. Runtimes start on the first message, are put away when idle with their session kept for the next message, and a small cap on concurrent runtimes protects the machine; when every slot is busy, the app holds your message and retries.

Borrowed, bounded authority. A runtime works as you, for this conversation, in this brand, and no wider. Before a turn runs, the bridge checks that you own the private conversation or are a live participant of the room. The runtime sees the brand documents it may read, its own working area and a read-only copy of the shared toolkit — never other people’s conversations, other brands, or any secret. Provider keys are never handed to a runtime: it calls a gateway with its short-lived credential, and the gateway checks the scope, adds the real key, makes the call and logs it against the run. Memory works the same way: every read and write carries the person and conversation it belongs to, checked in trusted code rather than by a prompt instruction. Isolation is per conversation rather than per person because two private conversations of one person must not read each other’s transcripts; if isolation cannot be verified at start-up, non-admin turns are refused.

Two loaders. Fixed identity is read as documents; live state is assembled by a program. They are separate because a document loader cannot query or filter, and because an instruction to “read this file” never fires on a surface where the agent is handed a conversation without a start-up turn. The program is the one loader that runs on every surface — chat, scheduled run and terminal — so all three start from the same knowledge. With no person present it returns the plate only, and says so.

The model is an engine, not the system. WilsonOS is a thin business control plane: it owns the brands, records, review path, memory and guardrails, and treats the model runtime as a replaceable harness. A model session is an execution detail, not a conversation’s identity.

Turns stay open while the agent works. A deliverable can take many minutes. Output produced outside an open turn is discarded, so agents finish their work inside the turn rather than handing it to a background helper that reports later.

Sources: docs/architecture/agents-runtime.md §The roster, §Boot, §Crons, §How agents run, §Per-conversation runtimes and isolation, §The scoped-execution contract, docs/architecture/memory.md §Boot loading — the two load paths; §P2 scoped memory, docs/architecture/chat.md (turn lifecycle), docs/architecture/channel-engine.md, docs/DECISIONS.md (ADR-022, ADR-028), agents/mia/CLAUDE.md and cron-registry.json (role, routines counted 2026-09-16), Don's product direction, 2026-09-16 (task 410 review, §E) and 2026-10-06 (agent table and character card; a COO per business), web/src/lib/brand-labels.ts (agent titles); web/public/images/agents/ (avatars) · Last verified 2026-10-06