The hub

Every team has a front door: a run that names no agent goes to the team's hub. It is the built-in orchestrator unless the team chooses one of its own agents, and either way it can plan, find the agent that keeps what a task needs, and hand it the work.

oa runs start platform 'Reconcile payment REF-9 against our open invoices.'

The built-in orchestrator¶

A team with no hub of its own gets the orchestrator. It runs on the team's model, so it cannot run until one is set:

oa teams settings set platform --model anthropic/claude-sonnet-5-5 --max-tokens 2048

Its tools:

Tool	What it does
`plan`	Keeps a short plan for the run and updates it as steps finish
`graph`	Searches the team's graph for agents, tools and skills
`delegate`	Hands a task to an agent it found, as a child run of its own
`web_search`	Searches the web through the team's provider — see Web tools
`web_fetch`	Reads one page, as markdown — see Web tools

Agents are found through the graph and listed nowhere else, so a team whose graph has no embedding model gets a hub with no delegate: it is told the team cannot be searched, and answers alone rather than guessing names.

Choosing your own hub¶

Any agent can be the hub. As the hub it has delegate, plan and the web tools whatever its own settings say; run as itself, it has only what it was given.

oa teams settings set platform --hub triage
oa teams settings set platform --no-hub      # back to the orchestrator

Deleting the agent that is the hub returns the team to the orchestrator.

What a delegate gets¶

A child run is built from its own agent: its model, its pattern, its grants. The hub's reach is not passed down, so handing work to an agent never gives it a server, a skill or the web it was not granted. The trace of the parent shows each delegation, and the child's own events are a run of their own:

oa runs events platform <run>

Not built¶

A new agent cannot be found by the hub for up to a minute. Agents reach the graph through a drain that ticks once a minute.
oa runs start <team> <agent> with no input asks the hub, treating the agent's name as the input.
Saving settings without the hub clears it. PUT /teams/{id}/settings is whole, so a client that leaves hub out returns the team to the orchestrator; oa teams settings set reads first and keeps it.
Renaming the hub agent can fail the next settings save, because settings name the hub by name.