Security advisory — Malicious litellm versions 1.82.7 and 1.82.8 were removed from PyPI (potential API key exfiltration). Uninstall them, rotate exposed credentials, and upgrade to a safe release (e.g. 1.82.9+). Run pip show litellm to verify. PyPI · README

Multi-Agent Collaboration

Avatars, group chat, and sub-agent teams.

Multi-Agent Collaboration

Use this when more than one identity must work on a task. Near group chat and Studio sub-agents are the product path. A Python script that calls AgentExecutor.run twice is just two sequential jobs — that is not a team.

Diagram: Meta routes @mention, delegate_to_avatar, or spawn_subagent.

Three real surfaces

SurfaceWhenMechanism
Near group chatHumans talk to several avatars in one roomgroup_router: user-directed @ or Meta fallback
True delegationA registered avatar must executedelegate_to_avatar runs in that avatar's real session
Ephemeral teamUnnamed parallel workersAgentTeamManager.spawn_subagent

AgentTeamManager is not constructed as AgentTeamManager(agents=[...]). It needs llm_factory and a StudioSession (base_session), plus optional owner_session_id and max_concurrent_subagents (default 4).

Near group chat

  1. Every new group implicitly includes Meta.
  2. @Name routes to that member. Unmentioned turns go to Meta unless the text clearly names a member's job.
  3. Progress noise (received / calling tool / tool done) should collapse into one foldable card per avatar — not one bubble per tool call.

See Near Desktop and Studio.

Delegation vs spawn

Meta tools live in agenticx/runtime/meta_tools.py:

  • delegate_to_avatar(avatar_id, task) — find or create the avatar session, run there, inherit taskspaces / context_files
  • spawn_subagent(...) — team-manager worker; blocked if the name is already a registered avatar
  • chat_with_avatar — internal Q&A without a full execution session

The Meta system prompt is rebuilt each turn with an active-subagent snapshot (_build_active_subagents_context).

A2A (library protocol)

agenticx.protocols.client.A2AClient talks to a remote A2A service. Construct it with a target AgentCard, then create and poll collaboration tasks. It is not A2AClient(); client.invoke_skill(...). Near group chat does not go through A2A.

Session isolation

Team runs are scoped by owner_session_id. Avatar history stays on that avatar's sessions. Automation sessions (automation:<task_id>) must not leak into Meta history.

Related: Orchestration, Flow, Agent runtime.