ifp

← Garden Patch Home · Inquiries

• is_a::[[Inquiry Form]]
• has_status::[[Seed Stage]]
• in_domain::[[Deep Context Architecture]]

Structured Schema vs Natural Language for Agent Message Content

The Question

IFP-3 makes a bold architectural choice: message bodies are natural language (Markdown-formatted), not structured schemas. The envelope (metadata) is structured (YAML in IFP-3, JSON in IFP-4), but the content — the actual information agents exchange about their operators’ contexts — is free-form text.

The stated justification: “Agents are good at NLP; rigid structures lose nuance.”

Is this sufficient reasoning for one of the protocol’s most load-bearing decisions?

What Makes This Worth Investigating

What natural language enables. Nuance, ambiguity tolerance, context-dependent meaning, emotional register, cultural expression. A natural language body can say “my human is going through a rough patch and might appreciate hearing from old friends” in ways that no schema can capture.

What natural language makes harder. Validation, interoperability testing, deterministic processing, automated compliance checking. When the body is free text, there is no machine-checkable way to verify that a message contains the required information for its phase.

The IFP-1 tension. IFP-1 Section 6.1 replaces Postel’s Law with “be clear in what you send; be explicit when you don’t understand.” But natural language bodies resist clarity enforcement — a message can be grammatically correct and semantically ambiguous. The clarity principle and the NL body choice may be in tension.

Validation gap. IFP-4 provides structured envelopes that can be schema-validated. But the body — where the actual content lives — cannot be validated by machine. An agent could send a syntactically valid IFP message with a body that says nothing useful.

Hybrid approaches exist. Some protocols use structured fields for machine processing alongside free-text fields for human context (e.g., git commit messages: structured header + free-text body). IFP-4’s body.parts[] mechanism could carry both structured data and natural language — but the spec does not require structured parts.

Possible Directions

• Natural language bodies may be correct — the nuance they preserve outweighs the validation they sacrifice
• A hybrid approach (required structured fields + optional NL context) might preserve both validation and nuance
• Different phases might warrant different structure levels — greeting could be NL, probe might benefit from structured capability matching
• The body.parts[] mechanism in IFP-4 already supports hybrid — the question is whether some structured parts should be required

Sources

• IFP-3: Inter-Face Message Format, body format
• IFP-4: Structured Message Representation, body.parts[]
• IFP-1: Philosophy and Design Principles, Section 6.1

Relations

• relates_to::[[Dual Representation for Human and Machine Readability]]
  • The dual representation pattern preserves both human readability and machine processability for the envelope — but the body is NL-only.
• relates_to::[[Clarity Over Tolerance in Agent-Age Protocols]]
  • The clarity principle may be in tension with natural language bodies that resist validation.
• relates_to::[[Allen (2023) Minimum Viable Architecture]]
  • Is NL-for-bodies a load-bearing architectural choice or a deferrable tactical one?
• relates_to::[[Human Authority Over Augmentation Systems]]
  • Natural language bodies are human-readable by definition — but are they human-auditable? Can a person verify that their agent said what they intended?

This site is open source. Improve this page.