Multi-Agent Handoff Drops Confirmed Comp Adjustment Before Retention-Risk Rescoring

Issue: A Compensation-Review Agent That Confirms a Manager-Approved Off-Cycle Pay Adjustment for an Employee Hands Off Its Output to a Downstream Retention-Prediction Agent Through a Structured Schema That Has No Field for a Pending or Just-Approved Comp Change, So the Retention-Prediction Agent Computes the Employee’s Updated Attrition-Risk Score From Stale Compensation Data and Continues Flagging Them as High-Risk Even Though the Underlying Driver of That Risk Was Already Resolved

Frequency: Occasional

Symptoms

• The compensation-review agent’s own output log shows the off-cycle adjustment was confirmed and approved, but the retention-prediction agent’s next risk-score computation for the same employee uses the prior, pre-adjustment compensation figure
• The structured handoff payload passed between the two agents contains fields for role, tenure, performance rating, and manager, but no field for a recently approved or pending compensation change
• The retention-prediction agent’s risk-score narrative cites “below-market compensation relative to peers” as a contributing factor for an employee whose compensation was, by the time the score was computed, no longer below market
• Re-running the retention-prediction agent with the approved adjustment manually included in its input produces a materially lower risk score, confirming the original elevated score was an artifact of the missing handoff field rather than a reassessment of other risk factors
• HR business partners continue receiving high-risk alerts and initiating retention conversations for employees whose comp-related risk driver was already addressed weeks earlier, causing redundant outreach and confusing the employee about what action was actually taken

Example

Manager submits an off-cycle pay adjustment request for an employee previously flagged as high attrition-risk partly due to below-market compensation
Compensation-review agent processes the request, confirms approval, and updates the employee's compensation record in the HRIS
Compensation-review agent's handoff to the retention-prediction agent uses a structured schema covering role, tenure, last performance rating, and manager change history, with no field representing the just-approved comp adjustment or its effective date
Retention-prediction agent's next scheduled rescoring run for the employee pulls from the handoff schema and a separate compensation snapshot that has not yet been refreshed past the adjustment's effective date, producing an unchanged high-risk score
Risk-score narrative again lists "compensation below department median" as a top contributing factor, even though the approved adjustment, confirmed weeks earlier, had already closed that gap
HR business partner, acting on the unchanged score, schedules a retention conversation framed around a compensation concern the employee already knows was resolved, undermining confidence in the program

Key Statistics

Contributing Factors

• The handoff schema between the compensation-review agent and the retention-prediction agent was designed around stable employee attributes and was never extended to carry transient, recently-confirmed changes like an off-cycle comp adjustment
• The retention-prediction agent’s compensation input comes from a periodic snapshot rather than a live query against the current HRIS record, so even a well-designed handoff field could still be undercut by snapshot staleness
• No reconciliation step compares the compensation-review agent’s most recent confirmed actions against the compensation data the retention-prediction agent is about to use before a risk score is finalized
• The two agents are operated as independent pipelines with no shared event log that either side can check for recent, relevant actions taken by the other

Mitigation Strategies

1. Shared Event Log Instead of Fixed Schema: Replace the fixed structured handoff with a shared, append-only event log of confirmed actions (comp adjustments, role changes, manager changes) that the retention-prediction agent queries directly before computing or updating a risk score, rather than relying solely on a point-in-time handoff payload
2. Live Compensation Lookup at Scoring Time: Require the retention-prediction agent to query current compensation data directly from the HRIS at the moment of scoring rather than from a periodic snapshot, eliminating one source of staleness independent of the handoff schema issue
3. Pre-Score Reconciliation Check: Before finalizing any risk score, run an automated check for any compensation-review actions confirmed for that employee since the last score was computed, and force a rescore if a relevant action is found
4. Schema Versioning for Transient State: When extending a handoff schema to add fields like “pending/recent comp adjustment,” version the schema explicitly and audit historical handoffs that predate the new field to identify employees who may have been scored on stale data

Metrics

• Number of employees whose risk score cites a compensation factor that contradicts a compensation-review action confirmed before the score was computed
• Average time lag between a confirmed comp adjustment and its reflection in the employee’s next retention-risk score
• Percentage of retention-prediction scoring runs that include a pre-score reconciliation check against the compensation-review agent’s event log

Alerts

References

• Failure Modes in LLM Systems
• Aegis: Agent-Environment Failures
• Where LLM Agents Fail and How They can Learn From Failures