Chapter 15: Structured Output and Schema-Constrained Generation (Claude Code Vs. Hermes Agent)

Key Takeaways

• Claude Code injects a synthetic tool named StructuredOutput to enforce JSON schema
• Retry budget defaults to five attempts, configurable via env variable
• Hermes forces a single function‑call tool, making the pattern provider‑agnostic
• Hermes validates JSON in‑conversation, feeding errors back as messages
• BatchRunner writes fully typed JSONL records, ensuring downstream schema consistency

Pulse Analysis

Structured output has become a cornerstone for production AI systems because it bridges the gap between natural‑language reasoning and the strict data contracts required by databases, APIs, and analytics pipelines. Claude Code’s implementation leverages Anthropic’s tool‑calling infrastructure by wrapping a user‑supplied JSON schema in a synthetic tool called StructuredOutput. This clever abstraction forces the model to emit a single, schema‑compliant call, while a built‑in retry budget caps the number of validation attempts, preventing runaway loops and surfacing clear error types for SDK consumers.

Hermes Agent takes a more universal route, exposing structured output as a model capability flag and using a single function‑calling tool that works across providers such as OpenAI, Anthropic, and local models. Validation occurs after the model’s response, with parsing and schema errors injected back into the conversation as user messages, giving the model contextual cues for self‑correction. This approach simplifies cross‑provider deployments and aligns with existing function‑calling APIs, while still offering configurable retry limits.

Beyond individual queries, Hermes scales structured extraction through its batch runner, which processes thousands of prompts in parallel and writes each result as a typed JSONL line. By normalizing tool statistics and trajectory metadata, the batch pipeline guarantees a uniform schema that downstream data warehouses and machine‑learning datasets can ingest without additional transformation. Together, these patterns illustrate how structured output is evolving from a niche feature to an essential engineering practice for reliable, enterprise‑grade AI applications.