Orchestration

This page explains how a Team leader coordinates members internally and the parameters that shape that behavior. All details are taken from buddy/team/team.py.

The leader delegates through tools

A team's leader is an LLM driven by the team's model. Rather than calling members directly in Python, the leader is given member tools: it invokes a member by emitting a tool call, the member runs, and its output returns to the leader as the tool result. The leader then either returns that result (route) or synthesizes a final answer from one or more member results (coordinate).

from buddy import Agent, Team
from buddy.models.openai import OpenAIChat

team = Team(
    members=[researcher, writer],
    mode="coordinate",
    model=OpenAIChat(id="gpt-4o"),   # the model that drives delegation
)

By default add_member_tools_to_system_message=True, so the leader is told which members and member-tools are available when it plans the run.

Mode dispatch

At run time the leader branches on self.mode:

• route — choose the single most appropriate member and return its output.
• coordinate — delegate subtasks to one or more members and merge the results into a single response.
• collaborate — have members work jointly toward a shared result.

Parameters that shape orchestration

Parameter	Default	Effect
enable_agentic_context	False	The leader can maintain a team context and automatically share it with members across the run.
share_member_interactions	False	Previous member interactions are forwarded to subsequent members, so they see what others produced.
get_member_information_tool	False	Adds a tool the leader can call to look up information about the team's members.
add_member_tools_to_system_message	True	Lists member tools in the leader's system message.
search_knowledge	True	If the team has knowledge, the leader gets a knowledge-search tool (Agentic RAG at the team level).

team = Team(
    members=[planner, coder, critic],
    mode="coordinate",
    model=OpenAIChat(id="gpt-4o"),
    enable_agentic_context=True,      # leader builds & shares working context
    share_member_interactions=True,   # members see each other's outputs
    get_member_information_tool=True,  # leader can query member capabilities
)

Why these matter

• enable_agentic_context is the difference between members working in isolation and the leader curating a shared, evolving brief. Turn it on when later members depend on earlier conclusions.
• share_member_interactions is lighter weight: it simply relays prior member outputs without the leader actively curating them.
• get_member_information_tool helps the leader route correctly when there are many members with similar-sounding roles.

Synthesis happens in the leader

In coordinate mode the final answer is produced by the leader's model, not concatenated from members. The quality of synthesis therefore depends on the leader's model and instructions.

Reasoning (optional)

The leader can run a reasoning pass before delegating. Relevant constructor parameters include reasoning (default False), reasoning_model, reasoning_agent, reasoning_min_steps (default 1) and reasoning_max_steps (default 10).

See Communication for how tasks and results actually flow between the leader and members.