Structured output

Pass a Pydantic model as output=. The runtime does its normal work, and its final answer is validated into result.output.

from typing import Literal
from pydantic import BaseModel

class Review(BaseModel):
    verdict: Literal["approve", "request_changes"]
    issues: list[str]

r = litellm.agent(
    Harness.CLAUDE_CODE,
    "Review the staged diff.",
    output=Review,
    permissions="read-only",
    sandbox=box,
    model="litellm_proxy/coder",
)
r.output.verdict  # "request_changes"

Harness	How the schema is enforced
CLAUDE_CODE	instruction to answer with one JSON object matching the schema, then validated
CODEX	Codex's native --output-schema
OPENCODE	instruction to answer with one JSON object matching the schema, then validated
DEEPAGENTS	create_deep_agent(response_format=Review)

If validation fails, the call raises OutputInvalid. The exception carries .raw with the model's text and .result with the rest of the turn, so the work isn't lost.