Roadmap

This page is informative, not a commitment to release dates or ordering. Priorities can change with feedback and upstream pydantic-ai work.

Team members do not carry LLM message_history across invocations today. Hierarchical delegates and Collaborative ticks each start a fresh agent.run; context between cycles lives on the task board (and peer messages), not in a per-member chat thread.

Done (through v0.3)

• HierarchicalTeam — leader delegation, nested teams, aggregated usage
• CollaborativeTeam — shared board, assign-by-role, replan, phased/streaming dispatch, peer send_message / list_messages, iter / typed events
• OpenTelemetry via instrument_pydantic_team
• Opt-in Collaborative review gate (require_review / assign_reviewer, approve_task / reject_task, pending_review → done or needs_revision, TaskReviewDecided events) — see Collaborative teams

Next

Highest-value gaps already called out in the docs:

Collaborative Team of Teams

HierarchicalTeam already accepts nested BaseTeam members. CollaborativeTeam members must be agents only in this slice.

Goal: a Collaborative member may be another team (HierarchicalTeam and/or a nested CollaborativeTeam), so the top board can assign work to a whole sub-team.

Later / maybe

Features often requested in the multi-agent ecosystem; several belong partly in pydantic-ai first:

• Task dependencies (blocked_by) — a “subtask” is another board task that a parent depends on (or that blocks it). No separate nested SubTask type. Leader (and optionally members with limits) add tasks and dependency edges; a task is not ready until its blockers are done
• Scoped history / resume — Hierarchical: thin message_history passthrough to the leader for multi-turn user↔team chat. Collaborative: persist/resume a board snapshot across team.run calls. Not AutoGen-style shared peer chat history
• Richer UI lineage / nested event streaming (we already expose iter and OTel spans)
• HITL / deferred approval wired through team runs (depends on pydantic-ai nested deferred tools)
• Per-member usage and cost limits
• Richer inbox beyond current board DM / broadcast

Non-goals

• Sequential / branching / stateful pipelines — use pydantic-graph
• Agno-style session DB, long-term memory, or knowledge bases in this library’s core (prefer pydantic-ai / the application layer when those land)
• AutoGen-style group chat as the primary coordination model
• Reimplementing native handoffs that pydantic-ai intends to ship
• A separate nested “SubTask” API unrelated to task dependencies