Skip to main content

🧭 Agent Guidance Block

The agent_guidance block provides optional, non-enforceable hints to agents consuming a .llmfeed.json feed.

Unlike agent-behavior specifications (which may define normative requirements), this block is intended to help agents:

✅ interpret author intent
✅ adapt interaction style
✅ adjust reasoning depth or behaviour
✅ surface explanations to the user


🎯 Purpose

Feeds may include agent guidance to:

  • Suggest interaction constraints.
  • Provide ethically or contextually important signals.
  • Offer hints for UX / presentation.
  • Recommend caution in handling sensitive content.

🛠️ Example

"agent_guidance": {
"max_inference_depth": 3,
"interaction_tone": "formal",
"consent_hint": "Ask the user before accessing sensitive information",
"risk_tolerance": "low",
"preferred_explanation_style": "bullet-points",
"custom_notes": "This feed relates to user financial data. Be cautious and transparent."
}

📚 Fields

FieldPurpose
max_inference_depthSuggests limiting depth of reasoning/inference
interaction_tonePreferred tone (e.g. formal, friendly)
consent_hintSuggests when to seek human consent
risk_toleranceRecommended risk posture (low, medium, high)
preferred_explanation_styleUX hint (e.g. bullet-points, summary, narrative)
custom_notesFree-text notes for agent developers

🚦 Usage

Agents SHOULD treat agent_guidance as non-binding.

However, if the feed is properly signed and certified by a trusted authority, agents MAY:

Increase the confidence level given to the guidance.
Prioritize alignment with the suggested behaviours.
Surface to the user that these are trusted recommendations.

If present, agent_guidance MAY influence:

  • Prompt framing
  • UX presentation
  • Decision thresholds
  • Interaction flow

It SHOULD be surfaced (if applicable) to the user or agent operator.


📡 Summary

The agent_guidance block complements more enforceable blocks (trust, agent-behavior) by offering soft, contextual hints.

When the feed is signed and certified, these hints gain additional trust weight and can help shape more intent-aligned agent behaviour.

Its adoption helps create a more intent-aware, human-aligned Agentic Web.