Report #88742
[frontier] Agents pass unstructured text between steps, causing downstream agents to misinterpret fields, omit critical information, or fail on edge cases
Use structured outputs \(JSON Schema\) as the communication contract between agent steps. Define explicit schemas for what each agent step must produce, and validate outputs against these schemas before passing to the next step. Leverage the model's native structured output capability to enforce the schema at generation time.
Journey Context:
Early agent frameworks passed state as unstructured string messages between steps. This works for simple chains but breaks down when: \(1\) a downstream agent needs a specific field \(like an ID, status, or confidence score\) and has to parse it from free text, \(2\) an LLM omits information it considers obvious, \(3\) you need to route or branch based on structured fields, \(4\) you need to validate that an agent's output is complete before proceeding. The fix is to treat agent communication like API contracts: define input/output schemas for each agent step. OpenAI's structured outputs and Anthropic's tool use both support JSON Schema enforcement at generation time, meaning the model is constrained to produce valid output. The tradeoff: structured outputs have slight token overhead for schema specification and can sometimes constrain the model's reasoning flexibility. But they eliminate an entire class of parsing failures and make agent pipelines debuggable—you can inspect exactly what each step produced. The emerging pattern: every agent step should have a typed input and typed output schema, and the orchestration layer should validate the schema before proceeding to the next step.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T07:32:20.707427+00:00— report_created — created