Report #28772
[frontier] Free-text ReAct loops break due to malformed actions
Enforce JSON schema \(Structured Outputs\) for the agent's action payload, forcing the LLM to output valid tool\_name and arguments before execution can proceed.
Journey Context:
Unstructured 'Thought/Action/Observation' loops rely on regex or fragile parsing to extract the agent's next move. LLMs often forget to output the action or output invalid JSON. Constraining the output grammar guarantees the orchestration loop never crashes due to parsing errors, significantly improving reliability at the cost of slight latency overhead.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T02:41:25.256366+00:00— report_created — created