Stop Parsing Agent Output Yourself: Get Validated JSON From Any Agent Workflow With the Claude Agent SDK

TL;DR: The Claude Agent SDK's outputFormat option validates agent output against a Zod or Pydantic model, giving you typed JSON with auto-retry — no manual parsing needed.

Every multi-step agent eventually returns something you need to parse. Ad-hoc json.loads() and regex string parsing works until the model rephrases its answer — then your pipeline breaks silently. The Claude Agent SDK has a cleaner primitive: outputFormat.

Define the shape once. The SDK validates the agent's output, re-prompts on mismatch (within a retry budget), and returns a fully typed object.

TypeScript / Zod:

import { query } from "@anthropic-ai/claude-code";
import { z } from "zod";

const BugReport = z.object({
  file: z.string(),
  line: z.number(),
  severity: z.enum(["critical", "high", "medium", "low"]),
  description: z.string(),
  suggestion: z.string(),
});

const result = await query({
  prompt: "Find bugs in src/auth/token.ts. Be thorough.",
  outputFormat: z.object({ bugs: z.array(BugReport) }),
});

// result.structured_output is fully typed — no cast needed
const critical = result.structured_output.bugs.filter(
  b => b.severity === "critical"
);

Python / Pydantic:

from anthropic_agent_sdk import query
from pydantic import BaseModel

class BugReport(BaseModel):
    file: str
    line: int
    severity: str
    description: str
    suggestion: str

class BugList(BaseModel):
    bugs: list[BugReport]

result = query(prompt="Find bugs in src/auth/token.ts.", output_format=BugList)
critical = [b for b in result.structured_output.bugs if b.severity == "critical"]

The SDK injects schema validation through the agent's StructuredOutput tool, so the model is constrained at generation time — not just post-checked. On validation failure it retries; after the retry budget you get a typed error, not silent garbage.

Use this wherever you'd otherwise parse agent output: code review pipelines, changelog drafters, dependency scanners, any workflow that fans out to structured data.

Sources: Get structured output from agents — Claude Code Docs · Structured outputs — Claude Platform Docs