Multi-Agent Handoff Drops Customs-Hold Flag Before Customer ETA Commitment

Issue: A Routing Agent Planning a Cross-Border Shipment’s Path Notes in Its Free-Text Reasoning That the Route Carries an Elevated Customs-Hold Risk at a Specific Border Crossing, but This Note Is Never Written to a Structured Field the Downstream Customer-Notification Agent Reads, So the Customer Receives a Committed Delivery ETA That Does Not Account for the Known Hold Risk

Frequency: Occasional

Symptoms

• The routing agent’s plan for a specific shipment names an exact border crossing and explains, in prose, why hold risk is elevated there this quarter — but the shipment record’s structured fields carry only transit times per leg, with no risk field at all
• The customer-notification agent generates the delivery ETA by summing the structured transit-time fields across every leg; it has no step that opens or parses the routing agent’s free-text planning notes before committing a date
• When the committed ETA passes and the shipment is still held at the flagged crossing, the resulting escalation record, compared against the routing agent’s original commentary, shows the risk was identified before the ETA was ever sent to the customer
• The gap is worst on crossings with historically volatile hold times, precisely because those are the cases where the routing agent’s judgment is qualitative rather than backed by a fixed rule — a crossing with a hard, published hold-time SLA would at least have a rule available to encode into a structured field
• No reconciliation step anywhere in the pipeline compares risk language in routing commentary against the structured record actually consumed downstream, so the mismatch is never caught before a customer-facing failure surfaces it

Root Cause The shipment record schema connecting these two agents was built to answer “how long will each leg take,” not “what could go wrong on this leg” — it carries a transit-time field per route leg but no field for a qualitative risk note tied to a specific crossing point. When the routing agent’s planning step identifies a customs-hold risk, that conclusion lives only in the free-text commentary attached to the routing plan, a field the notification agent’s ETA-generation logic was never built to read; it sums the numeric transit-time fields and commits a date. Because the two agents are invoked separately, at different points in the shipment lifecycle, there is no shared execution context in which the notification agent would encounter the routing agent’s reasoning unless it were explicitly written to a field the notification agent actually consumes.

Example

Routing agent plans a cross-border shipment and notes in its commentary: "Route crosses at the Laredo border point, which has shown elevated customs-hold times this quarter due to increased inspection volume -- recommend a wider ETA buffer for this leg"
Routing agent's structured shipment record, consumed by the customer-notification agent, includes the planned transit time for each leg but no customs-risk field; the hold-risk note exists only in the free-text commentary
Customer-notification agent generates the customer-facing ETA by summing the structured transit-time fields, with no adjustment for the noted customs-hold risk
Shipment is held at the Laredo crossing for four additional days; the committed ETA passes, and the customer escalates a delivery they were told would arrive on time despite the risk having already been identified internally

Key Statistics

Contributing Factors

• Routing agent’s output contract permits a customs-hold risk note to exist only in free-text commentary without requiring a corresponding structured field on the shipment record
• Customer-notification agent is designed to generate the ETA from structured transit-time fields only, with no step that re-scans routing commentary for risk language before committing a date
• No automated reconciliation compares risk-relevant language in routing commentary against the structured shipment record actually passed to the customer-notification agent

Mitigation Strategies

1. Mandatory Structured Risk-Flag Field for Border Crossings: Require the routing agent to populate a structured customs_risk field (with crossing point and risk level) for any route segment crossing a border, and block the shipment record from being marked complete without it
2. Commentary-to-Field Reconciliation Check: Run an automated scan of routing commentary for risk-relevant keywords (“customs,” “hold,” “inspection”) and verify a corresponding structured field exists before the shipment proceeds to ETA commitment, flagging any mismatch
3. Risk-Adjusted ETA Buffer for Flagged Crossings: Require the customer-notification agent to apply a defined ETA buffer whenever the structured customs_risk field indicates elevated risk, rather than computing the ETA from transit-time fields alone
4. Customer Communication Update on Materialized Hold: When a shipment with a flagged customs-risk field actually enters a hold, require an automated customer notification update rather than allowing the original, unadjusted ETA to silently lapse

Metrics

• Rate of committed ETAs generated for shipments where routing commentary contained customs-risk language but the structured record carried no corresponding field
• Percentage of border-crossing shipments with a structured customs_risk field populated versus commentary-only
• Customer escalation rate for delayed shipments, segmented by whether the delay was previously flagged in routing commentary

Alerts

• Customer-facing ETA is committed for a shipment whose routing commentary flagged customs-hold risk with no corresponding structured field → P1
• A shipment held at customs has no proactive customer communication update sent within a defined window after the hold is detected → P2
• Reconciliation check finds a routing-commentary-to-structured-field mismatch rate above baseline for two consecutive periods → P3

References

• Demystifying the Lifecycle of Failures in Platform-Orchestrated Agentic Workflows
• Why Do Multi-Agent LLM Systems Fail? (MAST)
• CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions