PydanticAI

• Agent[Deps, Output] is generic over a dependency type and an output type; the type checker sees both.
• Dependency injection: deps (DB pool, HTTP client, user id) are passed to run() and reach tools and dynamic system prompts through RunContext[Deps].
• Tools via @agent.tool; argument schemas and validation come from the function signature; validation errors are returned to the model as retry prompts (ModelRetry).
• Output types: a Pydantic model or union enforces the final answer shape; output_validator functions can reject and ask the model to retry.
• Model-agnostic with per-provider Model classes, FallbackModel, and first-class Pydantic Logfire / OpenTelemetry instrumentation; a pydantic_graph module exists for explicit state machines.

• Python teams that already use Pydantic and FastAPI and want agents that feel like ordinary typed code.
• Tool-heavy single agents where argument validation and retries matter (Argument Validation).
• Testable systems: TestModel and FunctionModel let you unit-test agent logic without network calls.

• Fewer batteries: no built-in RAG, document loaders, or memory modules — you bring your own, which is either a feature or a gap.
• Multi-agent support is delegation via tools; there is no supervisor/handoff runtime, and graph support is lower-level than LangGraph.
• Retry-on-validation loops can quietly burn tokens if a tool schema is ambiguous; set retries deliberately.
• Still pre-2.0-style churn: result/output naming and streaming APIs have been renamed across releases.
• Python only; no TypeScript story for full-stack teams.

• You need durable checkpointed workflows with human interrupts out of the box.
• Your stack is TypeScript.
• You want a large integration catalogue (loaders, vector stores) rather than a runtime.

1from dataclasses import dataclass
2from pydantic import BaseModel, Field
3from pydantic_ai import Agent, RunContext, ModelRetry
4 
5@dataclass
6class Deps:
7    db: Database
8    user_id: str
9 
10class Recommendation(BaseModel):
11    product_id: str
12    reason: str = Field(max_length=200)
13 
14agent = Agent("anthropic:claude-sonnet-4-5", deps_type=Deps, output_type=Recommendation,   # model id is version-sensitive
15              system_prompt="Recommend one product the user has not bought.", retries=2)
16 
17@agent.tool
18async def purchase_history(ctx: RunContext[Deps], limit: int = 20) -> list[str]:
19    """Product ids the current user already bought."""
20    return await ctx.deps.db.purchases(ctx.deps.user_id, limit)
21 
22@agent.output_validator
23async def not_already_bought(ctx: RunContext[Deps], out: Recommendation) -> Recommendation:
24    if out.product_id in await ctx.deps.db.purchases(ctx.deps.user_id, 1000):
25        raise ModelRetry("That product was already purchased; pick another.")
26    return out
27 
28result = await agent.run("Something for winter hiking", deps=Deps(db, "u-42"))
29print(result.output)              # validated Recommendation