4 patterns for this goal
Offer-generation agents produce compensation figures from stale training-data benchmarks instead of querying live market data, anchor offers to mismatched leveling precedents selected via embedding similarity, and commit to negotiated terms in offer letters without verifying those terms were actually captured in the structured handoff from the recruiting coordinator. These failures cluster around two categories: stale knowledge overriding live tools (compensation data, leveling precedents), and multi-agent handoff loss (negotiated exceptions, visa-contingent arrangements). Offer generation errors are silent — the output looks like a plausible offer letter — and materially affect allocational fairness (disparate offer bands by inadvertently using wrong precedents) and employer compliance (offers that contradict what the candidate was verbally promised).
All four offer-generation patterns root in the same architectural choice: treating agent-generated answers and retrieved live data as having equivalent reliability. When an agent has immediate access to a live compensation-benchmarking tool or a leveling-framework rubric, but the underlying model also carries parametric knowledge of typical figures or precedents, both produce plausible answers with identical fluency. Without an enforced requirement that certain decision-relevant data (market rates, live leveling rubrics, compliance-gate status) must be sourced from live tools rather than parametric knowledge, the agent defaults to whichever path is easier or most generative — and the stale path is often more fluent because it reflects high-frequency pretraining patterns. Mitigation is structural: mandatory tool-call requirements for decision-critical data, live status re-checks immediately before consequential actions (offer sending, background-check gates), and structured handoff fields that force explicit representation of every negotiated term and compliance gate.
Require the offer-generation agent to log a tool-call entry for every market-rate figure cited in the offer rationale. If a market-rate claim has no corresponding live-tool call in the trace, treat the offer as unverified and regenerate it with the live tool enforced. Independently query the live benchmarking tool for the same role/level/location and compare; a material mismatch indicates the offer was anchored to stale data.
Embedding-similarity retrieval over role descriptions ranks candidates by textual overlap (title vocabulary, industry keywords), not by the structured criteria (career track, leveling framework, scope of role) that actually determine comparable compensation. A “Senior Engineer” from one framework can share near-identical description text with a “Senior Engineer” from a different framework while having entirely different banding criteria. Pre-filter retrieved precedents by leveling-framework and career-track identity before applying text similarity.
Only by reaching out proactively — the candidate will see the discrepancy and may use it as leverage or lose trust in the hiring team. Mitigation requires reconciliation before the letter is sent: scan the recruiting coordinator’s negotiation transcript for any term not represented in the structured offer parameters, flag mismatches as mandatory holds, and require human review before the letter is finalized.
Require a fresh, synchronous call to the background-check vendor’s status API immediately before any offer letter is marked for sending, not a cached or task-list-derived state. Treat the absence of a check item from the agent’s internal task list as not equivalent to “cleared” — a cleared status must come from the vendor’s own status field explicitly equaling the defined “clear” value.
Leveling-framework errors affect the computed band (wrong precedent selection, wrong framework applied). Negotiation-handoff errors affect offer terms (visa contingency, relocation timing) — the compensation may be correct, but the conditional terms are wrong or missing. Both are silent failures: the offer looks complete, but critical context is missing.
| Pattern | Mechanism |
|---|---|
| Stale Training-Corpus Comp Benchmarks Override Live Market Data | Agent answers market-rate questions from pretraining knowledge instead of calling live benchmarking tool, producing offers anchored to outdated figures |
| Embedding Retrieval Pulls Mismatched Job-Family Leveling Precedent for Offer Band | Precedent retrieval by text similarity surfaces a role from a different leveling framework, misaligning the computed offer band |
| Multi-Agent Handoff Drops Negotiated Visa-Tied Remote Exception Before Offer-Letter Generation | Recruiting coordinator negotiates a visa-contingent remote-work exception that exists only in free-text notes, not in structured offer parameters; offer letter omits it |
| Offer Letter Auto-Sent Without Rechecking Live Background-Check Gate Status | Offer sent based on background-check item disappearing from agent’s task list (not re-queried) rather than confirming actual “clear” status from vendor’s live API |
Total: 4 patterns
An Offer-Generation Agent's Embedding Search over Past Leveling Decisions Surfaces a Role with Lexically Similar Title and Description but a Different Career Track or Leveling Framework, and the Agent Anchors the New Offer's Band to That Mismatched Precedent
A Recruiting-Coordinator Agent's Free-Text Negotiation Notes Recording a Visa-Status-Contingent Remote-Work Exception Are Not Captured in the Structured Handoff Schema Passed to the Offer-Letter-Generation Agent, Which Issues a Standard In-Office Offer That Contradicts What the Candidate Was Told
An Offer-Generation Agent Configured to Send a Finalized Offer Letter Automatically Once a Candidate Clears a Conditional-Offer Gate Sends the Letter Based on the Background-Check Step No Longer Appearing in Its Own List of Outstanding Blockers, Without Re-Querying the Background-Check Vendor's API for the Current Status, and the Check Had Actually Moved to "Pending Dispute" Rather Than "Clear"
An Offer-Generation Agent Answers Market-Rate Compensation Questions from Salary Figures It Absorbed During Pretraining Rather Than Calling the Live Compensation-Benchmarking Tool It Has Available, Producing Offers Anchored to Outdated Market Data for Fast-Moving Roles