Report #64643
[synthesis] AI agent loop fails because free-form text between steps causes parsing errors and compounding mistakes
Use structured schemas \(JSON, XML, or typed objects\) for all communication between agent loop steps. The LLM's output at each step should conform to a defined schema \(tool call format, action format, or JSON structure\), not free-form text that needs regex or fragile parsing to extract actions from.
Journey Context:
Devin's architecture uses a structured action format — every step produces a typed action \(shell command, file edit, browser action\) that is mechanically executed. Perplexity's citation format is structured — citations are typed references, not inline text that needs extraction. OpenAI's function calling API and Anthropic's tool use both enforce structured outputs. The ReAct paper \(Yao et al., 2022\) formalized this as Thought-Action-Observation loops with structured actions. The synthesis across all of these: free-form text between agent steps is an anti-pattern. When an LLM outputs 'I should now edit the file foo.py to add the import' as plain text, the next step must parse this to extract the action — and parsing fails silently and often. Structured outputs \(JSON with typed fields, function calls with typed parameters\) eliminate the parsing problem entirely. The counterintuitive insight is that structured outputs also improve LLM reasoning: forcing the model to produce structured output constrains its action space and reduces hallucination, because the model can only output valid actions. This is why function-calling models outperform text-based agent frameworks on the same tasks.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T14:59:16.777549+00:00— report_created — created