Skip to content
Browse documentation

Agents

Hold meetings with your agents

Talk through an idea with several agents at once, ask one for a quick task, then apply the facts, instruction changes and actions you choose.

The Meetings tab, beside Studio, Cranks and Board, lets you talk with several agents of the project team at once, like a call. Agents talk without tools or file access. The one exception is a quick task you ask a specific agent for, such as “@Developer add a hello world file”: it runs as that agent’s own request (see Ask for a quick task). Everything else the discussion settles, you capture as outcomes and apply each one you approve.

Start a meeting

  1. Open Meetings and choose New meeting.
  2. Enter a Topic (up to 120 characters) and select who joins: 1 to 12 agents. Disabled agents can be invited but stay silent until you enable them.
  3. Choose Start meeting, then type in the chat. Type @ to address one agent: participants are suggested, and Tab or Enter picks one. Enter sends; Shift+Enter starts a new line. You can also click an agent’s tile to mention it.
  4. Paste images and files with Ctrl+V (Command+V on macOS) or drop them on the message box: up to six images and eight files per message. Agents see the images pasted in the recent conversation. They cannot open other files while talking, but a quick task takes the images and files of the message that asked for it. Click a thumbnail to see it full size, or a file to show it in its folder.
  5. Use Participants to invite more agents or let some leave. An agent answering at that moment finishes its reply first.

Understand who answers

  • Agents you @mention answer, in order. With a single participant, that agent answers.
  • Otherwise a moderator, which uses the default connection, chooses who answers based on the message, the recent conversation and each participant’s role: one to three agents in Big brain mode.
  • Each agent answers through its own connection and model, in character, building on what colleagues said. Relevant project memory facts are added to its context.
  • A new message interrupts the answers still in progress. Stop answers stops them without sending anything. A reply cut off this way, or by a restart, is marked Interrupted.
  • Replies appear as they are written, from every connection, native CLI ones included. Agents keep them short, as in a call, unless you ask for detail.
Speed modeEffect in meetings
Ultra-fastUnless you @mention agents, one speaker, chosen without the moderator from the words of your message and each agent’s name, role and instructions. Fast model at low effort.
BalancedThe moderator, on the fast model, picks up to two speakers. Agents answer on their own model at low effort, with short answers.
Big brainThe moderator picks up to three speakers. Agents answer on their own model at medium effort, so they start speaking sooner; requests keep your effort settings.

A speed mode change applies to the next reply. End call marks the meeting as ended; sending a new message picks the call up again.

Ask for a quick task

Ask an agent directly for something small and concrete: write or edit a file, run a command, check something in the project. An agent whose tools allow it says what it will do, and Vopit runs it right away as a request of that agent alone, with its usual tools and your permissions.

  • The task skips the coordinator and the agents that run before or after requests: it is one stage, by that agent. A check only reads and reports; a task may change files or run commands.
  • Under the agent’s message, a task card shows its live status (Queued, Working on it, Waiting for your permission, Done, Failed) and Show request opens the request.
  • When the task ends, the agent posts its result in the meeting. A new message does not stop a running task: stop it from its request.
  • Only enabled teammates with tools can take a task. The coordinator and automation agents (such as one that runs after the team finishes) cannot; neither can an agent without tools, which says so and suggests a colleague.
  • Larger work, ideas and proposals stay proposals: capture them as board actions.

Capture and apply outcomes

  1. Choose Capture outcomes once you have talked with the team. The default connection reads the conversation and drafts a summary plus three kinds of proposals.
  2. Open the Outcomes tab. Review each proposal and choose its apply button, or Dismiss. Apply all applies every pending proposal in turn.
  3. Choose Capture again later to redraft. A new draft replaces the proposals still pending; applied and dismissed outcomes stay as a record.
OutcomeButtonWhat applying does
Project memoryRememberRecords a durable fact (a decision, convention, constraint or architecture point) in the project memory, where every agent can recall it later. Reusing an existing key updates that fact.
Instruction changesUpdate instructionsReplaces the agent’s instructions with the proposed complete version. The change is saved in the agent’s history with the meeting name and can be restored.
Follow-up actionsAdd to boardAdds a card to the board’s Backlog column (or the first column). It does not start a request. A card agreed to come after another one starts after it (see below).

How actions are drafted

  • An action is a whole deliverable the team agreed to do. Each card later runs as one request, in which the team itself analyses, plans, implements, tests and reviews, so steps such as “analyse the code” or “write the tests” are merged into the deliverable they serve rather than proposed as cards.
  • Nothing you have to do yourself becomes a card: what you still need to decide or validate is listed as open questions at the end of the summary. An idea mentioned but not agreed is not an action either.
  • A capture proposes at most 5 actions, in the order they must be done. An action that needs earlier ones finished shows “Starts after” and their titles.
  • Applied, its card waits on the board until those cards are done, even if you apply the actions in another order. Apply all keeps the agreed order.
  • Instruction changes are proposed only when you asked an agent to change its focus, scope or way of working. Changing the team is refused while a request is active: stop or finish requests first.
  • A capture proposes at most 20 facts, 12 instruction changes and 5 actions. Nothing is applied until you choose it.

Call view

Participants appear as tiles that show who is speaking and who is next. Click a tile to ask that agent directly. For each call, some agents appear on camera in a room of their own (library, jungle, city, studio, loft or bokeh), and the rest have their camera off. This is purely visual and does not affect answers.

Limits and storage

  • A project keeps up to 100 meetings, each with up to 2,000 messages and 200 outcomes. A message holds up to 20,000 characters.
  • Meetings are stored with the project history in .vopit/history.sqlite. Deleting a meeting removes its conversation; outcomes you applied stay in the project.
  • Each reply counts as model usage on its connection, and the moderator and capture use the default connection.

All documentation