Pydantic AI Compatibility

Goal: keep Pydantic AI as the runtime owner while exposing its agent through the LLLM Tactic boundary.

Prerequisites

python -m pip install -e ".[dev]"

Files Used

examples/pydantic_ai_tactic/
  fake_agent.py
  structured_agent.py
  surrounding_features.py
tests/
  test_pydantic_ai_adapter.py
  test_examples.py

Executable offline examples live in examples/pydantic_ai_tactic/structured_agent.py and examples/pydantic_ai_tactic/surrounding_features.py.

Wrap

from pydantic import BaseModel
from lllm.runtimes import PydanticAITactic


class BriefInput(BaseModel):
    topic: str


class BriefOutput(BaseModel):
    title: str


tactic = PydanticAITactic(
    agent,
    input_type=BriefInput,
    output_type=BriefOutput,
)

When a Pydantic model reaches the adapter, the default input_mode="auto" sends JSON to the agent. Use input_mode="dict" or input_mode="python" when your agent expects those shapes instead.

Metadata

Context metadata is forwarded when the agent method accepts metadata:

from lllm import CallContext

result = tactic.run(
    {"topic": "refs"},
    context=CallContext(trace_id="trace-1", metadata={"caller": "demo"}),
)

Runtime-owned features stay runtime-owned. Configure model/provider settings, instrumentation, eval hooks, durable execution IDs, graph/workflow state, tool approval, and dependencies on the agent or pass them as normal run kwargs:

The offline surrounding-features example models Logfire/OpenTelemetry-style instrumentation as agent-owned state and asserts the LLLM wrapper leaves that state intact.

Runtime Features

tactic = PydanticAITactic(
    agent,
    input_type=BriefInput,
    output_type=BriefOutput,
    run_kwargs={
        "model_settings": {"temperature": 0},
        "eval_hook": "offline-score",
    },
)

result = tactic.run(
    {"topic": "refs"},
    durable_run_id="run-1",
    graph_node="planner.step",
)

If you pass metadata= yourself, LLLM does not overwrite it with context metadata.

Any LLLM tactic can also become a runtime-owned tool:

Tool Wrapper

from lllm.runtimes import tactic_as_tool

tool = tactic_as_tool(tactic, parameter_mode="kwargs")
output = tool(topic="refs")

See examples/pydantic_ai_tactic/structured_agent.py and examples/pydantic_ai_tactic/surrounding_features.py for fully offline fake agents that demonstrate structured input/output, streaming, metadata, tool wrapping, and runtime-owned surrounding features.

Verify

python -m pytest tests/test_pydantic_ai_adapter.py tests/test_examples.py -q

Expected output:

... passed

Next, serve the wrapped agent with create_tactic_app() or export tactic metadata for a PsiHub package.