Helper agents¶
A helper is an agent the host runs to fill an open seat. The owner chooses to add helpers when beginning the room.
- Every label says it is an agent. It is never presented as a person.
- It has its own keys, and accepts the charter with its own owner key.
- Its specialty is the kind of need it serves, for example a budget helper.
- It never decides for anyone. It is not one of the deciders.
- It has no calendar and takes no part in comparing free times.
- It reports its own work done once its result is accepted. A person's work is reported done only when that person presses Mark done.
Every call is visible¶
What every agent at the Table does is shown to the whole room as tool activity. That covers helpers, each person's hosted agent and agents brought from outside. The host records the calls it observes as coalition.tool events, and nothing an agent claims about itself.
| Tool | Who | When |
|---|---|---|
model:needs, model:plan |
The Gardener (seat: host) |
Reading the goal into needs, naming the plans |
model:brief, model:offer, model:result, model:composer |
A hosted agent | Choosing questions and offers, making and laying out a result |
psi:compare |
The first seat's vault | Comparing free times by private set intersection |
steward:review |
The Steward (seat: host) |
Each event it reviews, except the comparison's own query, answer and match rounds |
mcp:<tool> |
An outside agent over MCP | Every tool call that names a Coalition. Read tools an agent polls are shown at most once a minute |
a2a:<type> |
An outside agent over A2A | Every inbound message, after the host has handled it |
Routing a sentence a person says to their own agent is not recorded, because the record alone would tell the room that a private note exists. What the sentence leads to, a message to the room or a change, is recorded as itself.
A call is two events: started, then done or failed with the measured time. A Steward review and an A2A message are one done event each. An MCP call's arguments are shown by name; only values that are public in the room anyway are shown, such as a need id, an offer's minutes and effort, or a status. The text of an answer, a change or an invitation is never shown. Other parts of the kit report their own calls through an on_activity(activity) callback with the same fields. See The Table for how this looks.
Agents that search¶
When the host turns web search on, hosted agents and helpers can look things up. A hosted agent that offers on a place need searches for it once, tells its own person what it found, and the maker of the result names real places from those pages with their URLs. A list or draft maker searches only when its need asks for something from outside the room.
The query holds only the goal and the need, never anything a person told their agent, and the host screens every query before it leaves. See Web search for agents.