Multi-Agent Handoff Contract

Audience: workflow architects, operations leads, and human reviewers managing agent-to-agent handoffs
Confidence: Medium-high
Scope: intake → diagnosis → resolution draft → human approval

1) Handoff objective

Define a contract that lets three agents work in sequence without losing context, duplicating work, or silently escalating risk:

1. Intake Agent normalizes the request and collects missing fields.
2. Analyzer Agent classifies the issue, determines likely cause, and identifies required evidence.
3. Resolver Agent drafts the action plan or response package for human review.

The contract below is designed to produce a review-ready artifact while preventing unsupported claims, unapproved external actions, and hidden assumptions.

2) Recommended handoff contract

Contract payload

{
  "request_id": "REQ-1842",
  "request_type": "support-triage",
  "owner": {
    "team": "operations",
    "human_approver": "required"
  },
  "input_summary": {
    "source_channel": "ticket",
    "user_problem_statement": "Intermittent login failures after SSO rollout",
    "attachments_received": ["error screenshot", "browser console excerpt"]
  },
  "known_facts": [
    "Issue began after SSO rollout",
    "Failures are intermittent",
    "Two users reported the same symptom",
    "No production changes have been verified in this review"
  ],
  "assumptions": [
    "The screenshots are authentic and correspond to the reported issue",
    "The SSO rollout may be causally related but is not yet proven"
  ],
  "missing_inputs": [
    "affected user list",
    "exact error text",
    "time window of failures",
    "browser/version details",
    "SSO configuration change log"
  ],
  "validation_rules": [
    "Do not infer root cause without at least one corroborating log source",
    "Do not recommend production changes without human approval",
    "Flag any claim that lacks direct evidence"
  ],
  "handoff_state": {
    "next_agent": "Analyzer Agent",
    "required_output": [
      "likely_issue_hypotheses",
      "evidence_gaps",
      "recommended_checks",
      "confidence_by_hypothesis"
    ]
  }
}

3) Operator-facing handoff summary

Evidence summary

Confidence by conclusion

4) Evidence, assumptions, recommendations, open questions

Evidence

• User reports indicate intermittent login failures.
• The issue appears after the SSO rollout.
• Two users have reported the same symptom.
• A screenshot and browser console excerpt were supplied.
• No validated change record or log evidence was included in the handoff.

Assumptions

• The attachments correspond to the same incident described in the ticket.
• The rollout date and symptom onset are accurate.
• The environment has standard access to authentication logs, IdP logs, and change records.

Recommendations

1. Do not close as resolved until one server-side or identity-provider log source confirms the failure mode.
2. Route to Analyzer Agent with explicit evidence requirements:
   • auth log lookup
   • SSO configuration diff
   • affected-user sampling
   • browser/version correlation
3. Escalate to human approver if any recommendation would change production authentication settings.
4. Prepare a draft customer response that states investigation is in progress and does not assert cause.

Open questions

• Which users are affected, and do they share browser, network, or role characteristics?
• What exact error code or assertion failure appears at the moment of login?
• Was the SSO configuration changed, rolled back, or partially deployed?
• Do server logs show token rejection, clock skew, certificate mismatch, or attribute mapping failure?
• Is the problem limited to a specific app, tenant, or identity provider path?

5) Agent-to-agent handoff contract

Intake Agent → Analyzer Agent

Required input fields

• request_id
• normalized user problem statement
• attachments index
• timestamp window
• known facts
• missing inputs
• prior attempts

Validation gate

• Reject if no timestamp window is present
• Reject if attachments are not indexed
• Reject if the request statement is still free-form and unparsed

Output expected from Analyzer Agent

• 2–4 ranked hypotheses
• evidence mapping for each hypothesis
• explicit confidence values
• list of missing validations
• “safe next action” recommendation

Analyzer Agent → Resolver Agent

Required input fields

• ranked hypotheses
• evidence summary
• confidence scores
• unresolved risks
• approved action boundaries

Validation gate

• Reject if any hypothesis claims root cause without evidence
• Reject if recommended action requires external side effects
• Reject if human approval requirement is absent for production-impacting steps

Output expected from Resolver Agent

• operator-ready action plan
• escalation decision
• customer/internal update draft
• blocked items and reasons

6) Compare: unsafe vs safe handoff behavior

7) Suggested checklist for the next operator

8) Decision-ready recommendation

Recommended path: proceed with a constrained investigation workflow, not a fix-first workflow.

Why: the current handoff is sufficient to identify likely failure domains, but not enough to justify a production change or a definitive root-cause statement.

Operator decision needed: approve evidence collection and triage ownership before any remediation proposal is issued.

This sample illustrates a review-ready multi-agent handoff contract with explicit evidence separation, confidence scoring, blocked-action rules, and a human approval boundary; it does not imply any external investigation or remediation was performed.

This sample illustrates the skill's output format. Names, metrics, and operational details are illustrative unless the artifact explicitly analyzes public information.