Human Interaction Protocol

When agents get it wrong, who is accountable?

AI agents are making real decisions in real workflows. Mistakes are inevitable. Without a shared interaction standard, it’s unclear:

• Who is responsible for approving, correcting, or owning an outcome
• What exactly is being asked (decision vs approval vs clarification) and what “done” means
• How to audit what happened later (who said what, when, and why)

HumanInteraction is a practical protocol that standardizes how agents ask humans for input and how humans respond, so teams can collaborate with clarity and share accountability.

What we explore

• Teamwork: why effective agents require humans and AI working together as a team
• Accountability: how to track AI actions and tie outcomes back to responsible human decisions
• Human factors: why interaction design must acknowledge fatigue/overload and usability/security considerations (as NIST recommends)
• Examples: how a shared protocol enables smooth, transparent interactions across stakeholders and channels

Core concepts

HumanInteraction

The agent’s structured request for human involvement (who / what / why), including constraints and context.

HumanFeedback

The human response that resolves (or partially resolves) an interaction, linked for auditability.

Expectation vs commitment

Soft asks vs binding promises. Commitments require explicit acknowledgement and can carry due dates and consequences.

Resolution policy

How multi-party inputs resolve—e.g. first response, majority, or all required actors.

Coordination mode

Centralized vs distributed handling when several humans participate (with optional groupId linking).

Roles

Primary, approver, contributor, observer—plus private vs group visibility when messages fan out.

Example message

Structured JSON keeps LLM generation simple and tooling interoperable:

{
  "interactionId": "abc123",
  "groupId": "grp789",
  "coordinationMode": "centralized",
  "resolutionPolicy": "majority",
  "timestamp": "2025-06-24T10:00:00Z",
  "agentInstanceId": "agent-XZ01:exec-5589",
  "participants": [
    { "id": "user-108", "role": "compliance_analyst" },
    { "id": "user-222", "role": "finance_approver" }
  ],
  "agentOutput": {
    "summary": "This invoice exceeds the threshold.",
    "confidence": 0.62
  },
  "type": "approval",
  "urgency": "blocking",
  "userActions": ["approve", "edit", "reject", "submit_text"],
  "status": "awaiting_input",
  "humanFactors": {
    "trustLevel": 0.8,
    "fatigueScore": 0.2
  },
  "history": []
}

Status

This protocol and its design docs are an evolving draft—naming and schema details may change as requirements and examples converge. Read the design deep dive for scope, requirements, and edge cases.

---

Next: Protocol specification · Design deep dive · Contributing