Self-improving AI Employees for everyone
Orchestrator Agent (orchestrator_agent)
| Field | Value |
|---|---|
| Category | specialized_agents |
| Plugin | server/nodes/agent/orchestrator_agent/__init__.py -> _specialized.py::SpecializedAgentBase.execute_op (dispatch via BaseNode.execute()) |
| Team lead | yes -- input-teammates handle enabled (team_lead_agent_handles()) |
| Tests | server/tests/nodes/test_specialized_agents.py |
Purpose
Team-lead agent that coordinates agents wired through input-teammates.
Task Manager is intrinsically bound and is the only model-facing delegation
interface. Internal delegate descriptors resolve authorized assignees after a
durable assign_task record exists.
What is unique to this node
input-teammateshandle (extra input beyond the shared 5; declared viateam_lead_agent_handles()in_handles.py).- No extra parameters: uses the same
SpecializedAgentParamsas every other agent. There is noteamMode/maxConcurrentfield — those were removed (or never existed on the current Pydantic model). tool_descriptiondeclared so a parent team lead delegating to this node gets a meaningful tool string ("ONE-SHOT delegation ...").- Team lead detection:
prepare_agent_call(in_inline.py) checksnode_type in TEAM_LEAD_TYPES = {'orchestrator_agent', 'ai_employee'}and callscollect_teammate_connectionsto build internal teammate descriptors. - Task control panel: selecting the node opens its non-removable Task Manager middle panel.
Teammate collection
collect_teammate_connections(node_id, context, database) in
server/services/plugin/edge_walker.py:
- Scans
context.edgesforedge.target == node_idandedge.targetHandle == 'input-teammates'. - Resolves the source node in
context.nodes. - Filters to
node_type in AI_AGENT_TYPES. - Loads
database.get_node_parameters(source_id)for each teammate. - Returns a list of
{node_id, node_type, label, parameters}dicts.
prepare_agent_call walks each teammate's own input-tools edges to populate
capability descriptions. The LLM receives connected node IDs and calls
task_manager(operation="assign_task", ...); direct delegate tools are hidden.
Behaviour
Inputs, parameters, outputs, logic flow -- see Generic Specialized Agent Pattern. The only difference is the teammate expansion above.
Edge cases
- Non-
AI_AGENT_TYPESnodes wired toinput-teammatesare silently skipped. - Concurrency is root-scoped and defaults to three active descendants. Additional durable assignments remain queued.
- When zero teammates are connected, the node behaves identically to any other specialized agent.
Related
- Pattern doc:
_pattern.md - Sibling team lead:
aiEmployee - Architecture: Agent Teams