Skip to content

Conversations

Chat is where most work starts — from a laptop or the phone app — but it is not where work is stored. A conversation is a place to set goals, answer questions, redirect, and watch progress. The work itself lands in tasks and deliverables that the conversation links to.

Every person has one private conversation per agent by default — you and Scout, you and Prime, you and Mia — and can open more. A private conversation belongs to its owner: what is said and remembered in it is scoped to that person and that agent.

A shared conversation (a room) has explicit membership. A partner or a client can be added to see the same thread and talk to the same agent. Membership is the authorisation: if you are not a participant, the room does not exist for you, and no amount of brand access changes that. Human-only rooms, with no agent in them, are a proposed extension.

Long tasks are not black boxes. As the agent runs, the conversation shows a nested trail:

Running flow social-content-engine
└ Running skill hook-extract
└ Calling tool image generation

Flows call skills, skills call tools, and all three layers stream to you as they happen — so you can see which method it chose (promise 2 of the Five), catch a wrong turn early, and understand what a five-minute silence was spent on. When a paid step is coming, the agent quotes it and waits for your go-ahead before spending.

Anything the agent makes or refers to has a stable address, and the agent refers to it by that address rather than by a file path or a description. In the conversation those references render as cards and links:

The agent refers to… You see…
A generated image or video A media card with a thumbnail that opens in a viewer
A task A link into the task and its deliverables
A skill or a flow A link to what that capability is and how it has been rated
A brand, or a tab within it A link straight into the brand workspace

Because the reference is an identity and not a location, it keeps working when the file is renamed, moved, or reorganised. Four different ways of spelling a path each broke rendering silently before this rule existed; identities do not.

When an agent is one short question away from the right output — three options, a yes/no, a choice of direction — it asks with a real question form, not a paragraph you have to parse and retype. You tap an answer and the agent continues from where it paused. It reserves this for finite choices that change the work; open-ended prompts and confirmations it would proceed on anyway are not turned into forms.

This is promise 4 in practice: knows when to stop and ask. A decision that is yours — approving spend, publishing, changing a brand rule — is put to you, not guessed.

  • Not a task. A message can create or reference a task; the task lives in Work & review with its own history.
  • Not the agent’s memory. What you decide in conversation is captured into memory by explicit writers during and after the session — not by the chat log being re-read later. See What your agent remembers.
  • Not the only trigger. Schedules and signals from connected channels start work too; their results are delivered to a conversation you chose. See Schedules & triggers.

live Private conversations, streamed flow/skill/tool progress, smart links, and question forms are live. direction Shared conversations with explicit membership are deployed in code with acceptance pending; human-only rooms are proposed. See Release scope.

Sources: docs/architecture/chat.md (streaming protocol, smart links, turn lifecycle), docs/architecture/chat-thread-model.html, input/WilsonOS_Overview.html §02 · Last verified 2026-09-15