pydantic-team

Type-safe team orchestration for pydantic-ai Agents.

v1 provides hierarchical and collaborative teams for pydantic-ai Agents:

• HierarchicalTeam — leader delegates via tools (agent delegation)
• CollaborativeTeam — shared task board with parallel claim/assign (phased or streaming)

For sequential, branching, or stateful pipelines, use pydantic-graph (already pulled in by pydantic-ai-slim).

Install

uv add pydantic-team
# optional Logfire backend:
uv add 'pydantic-team[logfire]'
# or from a checkout:
uv sync --group lint --group dev

Depends on pydantic-ai-slim (core agents only — no provider SDKs). Install a provider extra when you need a live model, e.g. pydantic-ai-slim[openai].

Requires Python 3.10+.

Documentation

Published docs: etiqa.github.io/pydantic-team

Local build (MkDocs + mkdocstrings):

uv sync --group docs
make docs-serve   # http://127.0.0.1:8000
make docs         # build into site/

HierarchicalTeam

The leader registers each member as a tool and passes usage=ctx.usage on nested runs so TeamResult.usage includes every agent involved.

import asyncio

from pydantic_ai import Agent
from pydantic_team import HierarchicalTeam

researcher = Agent(
    'openai:gpt-4.1',
    name='researcher',
    instructions='Research the topic and return concise notes.',
)
writer = Agent(
    'openai:gpt-4.1',
    name='writer',
    instructions='Turn research notes into a short article.',
)

team = HierarchicalTeam(
    leader_model='openai:gpt-4.1',
    members=[researcher, writer],
    system_prompt_override=(
        'Delegate to researcher or writer based on the task, then synthesize a final answer.'
    ),
)


async def main() -> None:
    result = await team.run('Explain pydantic-ai agent delegation briefly.')
    print(result.data)
    print(result.usage)


asyncio.run(main())

You can also pass an existing leader_agent= instead of leader_model=. Nested HierarchicalTeam instances are valid members (set name= for a clear tool id).

See the hierarchical teams guide for nested teams, usage details, and TestModel testing.

CollaborativeTeam

Shared TaskBoard: the leader creates tasks and assigns them by role; members complete their assigned work in parallel rounds (max_rounds). Use dispatch_mode='streaming' so members start as soon as they are assigned (overlap with seed/replan); the default 'phased' keeps a seed barrier. Optionally cap how many assignments each member sees per tick (max_assignments_per_tick). If work remains, the leader can replan (max_replans, default 0) before the final synthesize. Each seed / replan / synthesize / member tick is an isolated agent run for usage limits; team usage aggregates them. Synthesize is toolless (no board mutation). Members may message each other directly (send_message / list_messages); the leader can observe with list_messages.

from pydantic_ai import Agent
from pydantic_team import CollaborativeTeam

researcher = Agent(
    'openai:gpt-4.1',
    name='researcher',
    instructions='Complete only research tasks assigned to you.',
)
writer = Agent(
    'openai:gpt-4.1',
    name='writer',
    instructions='Complete only writing tasks assigned to you.',
)

team = CollaborativeTeam(
    leader_model='openai:gpt-4.1',
    members=[researcher, writer],
    max_rounds=3,
    max_replans=2,
    dispatch_mode='streaming',
)
result = await team.run('Draft a short brief on agent teams')

Or observe the run step by step:

async with team.iter('Draft a short brief on agent teams') as run:
    async for event in run:
        ...  # TasksScheduled / TaskCompleted / PhaseJoined / MessagePosted / RunEnded
    assert run.result is not None

Result type

from pydantic_team import TeamResult

# result: TeamResult
# result.data  — final leader output
# result.usage — aggregated RunUsage (requests + tokens)

Sequential / complex control flow

Use pydantic-graph when you need an ordered pipeline, branches, loops, or shared state. This library intentionally does not reimplement that.

Examples

Runnable scripts (live model API — not part of the test suite). Core deps stay pydantic-ai-slim only; the examples group pulls in the OpenAI provider extra, logfire, and python-dotenv. Examples call logfire.configure(send_to_logfire='if-token-present'), logfire.instrument_pydantic_ai(), and instrument_pydantic_team() so spans print locally without auth; set LOGFIRE_TOKEN or run logfire auth for cloud.

# Put OPENAI_API_KEY in examples/.env (auto-loaded) or export it.
# optional: export PYDANTIC_TEAM_MODEL=openai:gpt-5.6-luna
uv sync --group examples
uv run examples/hierarchical_basic.py
uv run examples/collaborative_basic.py
# one-shot without a prior sync:
# uv run --group examples examples/hierarchical_basic.py

Development

make install      # uv sync + pre-commit
make format
make lint
make typecheck
make test         # pytest with --cov-fail-under=100
make ci           # format + lint + test
make ci-strict    # lint + typecheck + test
make docs         # MkDocs build (needs --group docs)
make docs-serve

Unit tests use pydantic-ai TestModel only — no live LLM calls.

On a v* tag, CI publishes to PyPI (Trusted Publisher / environment release) and deploys docs to GitHub Pages.

License

MIT