Inter-Agent Communication

This page describes how tasks and results move between a Team leader and its members, and how shared state lets members build on one another. Details are grounded in buddy/team/team.py.

The basic loop

1. You call team.run(message) (or print_response).
2. The leader receives the message and decides which member(s) to involve.
3. The leader delegates a subtask by calling a member tool; the member runs as an agent and returns a RunResponse.
4. The member's output comes back to the leader as the tool result.
5. The leader either returns it (route) or synthesizes a final TeamRunResponse (coordinate).

Members do not talk to each other directly — the leader mediates every exchange.

Seeing member responses

By default members work quietly behind the leader. Set show_members_responses=True to surface each member's contribution, which is invaluable while developing.

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

team = Team(
    members=[researcher, writer],
    mode="coordinate",
    model=OpenAIChat(id="gpt-4o"),
    show_members_responses=True,
)
team.print_response("Draft a release note for v2.2.0.")

Sharing context between members

Two parameters control how much one member sees of the others' work:

Parameter	Default	Effect
share_member_interactions	False	Forwards previous member interactions to later members
enable_agentic_context	False	Lets the leader maintain a team context and share it with members

team = Team(
    members=[planner, coder, critic],
    mode="coordinate",
    model=OpenAIChat(id="gpt-4o"),
    share_member_interactions=True,  # critic sees planner & coder output
    enable_agentic_context=True,     # leader curates shared context
)

Use these when later members depend on earlier results — for example a critic that must review what the coder produced.

Shared state

A team carries dictionaries that persist information across a run and across turns:

Field	Scope
session_state	Per-session state, persisted to storage when configured
team_session_state	State shared between the leader and members
workflow_session_state	State shared when the team runs inside a workflow

add_state_in_messages (default False) injects state variables into the user and system messages so the model can read them.

team = Team(
    members=[agent_a, agent_b],
    model=OpenAIChat(id="gpt-4o"),
    team_session_state={"customer_tier": "gold"},
    add_state_in_messages=True,
)

History

To let the leader remember earlier turns of a conversation, enable history:

• add_history_to_messages (default False) — include prior runs in the messages sent to the model.
• num_history_runs (default 3) — how many past runs to include.

team = Team(
    members=[agent_a, agent_b],
    model=OpenAIChat(id="gpt-4o"),
    add_history_to_messages=True,
    num_history_runs=5,
)

enable_team_history is deprecated

The older enable_team_history flag is deprecated in favor of add_history_to_messages. Prefer the latter in new code.

Results

team.run() returns a TeamRunResponse whose .content holds the leader's final synthesized answer. When stream=True, run() instead yields an iterator of run/team events you can consume incrementally.