Multi-Step Onboarding Agent Loses Context on Conditional Task Across Sessions
Issue: An Onboarding Agent That Generates a New Hire’s Task Checklist Establishes a Conditional Requirement Early in the Onboarding Conversation (e.g., “Since You’re an International Hire, You’ll Also Need to Complete X”) but Loses Track of That Condition in a Later Session, Generating a Follow-Up Checklist That Omits the Conditional Task Entirely
Frequency: Occasional
Symptoms
- A conditional onboarding task explicitly identified and communicated in an earlier session (visa-status paperwork, equipment-shipping for a remote international hire, a state-specific compliance form) is absent from a later checklist or reminder generated by the same onboarding workflow for the same hire
- The omission is traceable to the conditional fact (e.g., “international hire,” “remote in a new state”) having been established only within that earlier session’s conversational context, with no corresponding structured field written to the new hire’s persistent record
- New-session checklist generation, when the conditional fact is re-stated explicitly in the new session’s prompt, correctly includes the conditional task – isolating the failure to context not persisting between sessions rather than to the model’s ability to reason about the condition when it is present
- The pattern recurs disproportionately for hires whose onboarding spans multiple distinct agent sessions (e.g., HR-coordinator session, IT-provisioning session, payroll-setup session) rather than a single continuous conversation
- Compliance audits catch a backlog of incomplete conditional paperwork concentrated among hires onboarded through the multi-session flow, well after the relevant deadline (e.g., visa filing window) has passed
Root Cause Each onboarding-workflow session is a separately invoked agent call bounded by its own context window; a fact established conversationally in one session (the hire’s international status, their work state) is only available to a later session if it has been written into the new hire’s structured, persistent record rather than left in that session’s own transcript. Because establishing a conditional task and persisting the underlying fact that triggered it are two different actions, an agent can complete the first (communicating the requirement to the hire in the moment) without reliably completing the second (writing the triggering fact to a field a later session’s agent will actually read).
Example
HR-coordinator onboarding session for a new hire establishes, partway through, that the hire will be working remotely from a state the company has no prior payroll registration in, and tells the hire "you'll need to complete state tax withholding form X in addition to the standard package"
This fact -- remote work state requiring additional withholding paperwork -- is communicated in that session's chat but never written to a structured "work_state" or "additional_compliance_required" field on the hire's persistent onboarding record
Payroll-setup session, run as a separate agent invocation a week later, generates its checklist from the structured record alone, which has no flag for the additional withholding requirement, and omits it
Hire's first paycheck is processed without the required state withholding election on file, discovered only when the payroll system itself later flags a mismatch
Key Statistics
| Finding | Source |
|---|---|
| Multi-agent and multi-session LLM workflows exhibit measurable information loss at the boundary between sessions or agents, where a fact generated in one context fails to propagate into the structured state a later session consumes | Why Do Multi-Agent LLM Systems Fail? (MAST) |
| LLMs measurably lose track of earlier-established facts across multi-turn interactions when those facts are not re-grounded in persistent, structured state | LLMs Get Lost In Multi-Turn Conversation |
| Memory architectures for autonomous LLM agents are an active research area specifically because conversational context does not reliably persist across agent invocations without an explicit memory mechanism | Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers |
Contributing Factors
- Onboarding workflow’s session boundaries do not require every conditionally-triggering fact identified in conversation to be written to a structured, persistent field before the session ends
- Later-session agents are designed to consume the structured onboarding record only, with no step that re-scans prior sessions’ transcripts for conditional facts
- No automated reconciliation compares conditional tasks communicated to the hire in any session against the structured checklist fields actually populated for that hire
Mitigation Strategies
- Mandatory Structured Field for Every Conditional Trigger: Require any session that identifies a conditional onboarding requirement to write the triggering fact to a structured, persistent field on the hire’s record before the session can be marked complete, not merely to communicate it conversationally
- Cross-Session Checklist Reconciliation: Before a later-stage session (payroll, IT provisioning) finalizes its checklist, run an automated check against every conditional task or fact mentioned across all prior sessions’ transcripts for that hire, flagging any conditional item not reflected in the structured record
- Single Persistent Checklist Owner: Maintain one canonical, structured onboarding checklist per hire that every session reads from and writes to directly, rather than each session generating its own checklist from a potentially incomplete structured snapshot
- Deadline-Linked Conditional-Task Audit: For conditional tasks with a hard deadline (visa filing, tax election windows), run a dedicated automated audit ahead of the deadline across all in-progress onboarding records, independent of whichever session last touched the hire’s file
Metrics
- Rate of conditional onboarding tasks communicated in one session but absent from a later session’s generated checklist for the same hire
- Time between a conditional task’s deadline and its detection as incomplete, segmented by single-session vs. multi-session onboarding flows
- Percentage of conditional-trigger facts identified in conversation that are written to a structured field within the same session
Alerts
- A conditional onboarding task with a compliance deadline is found incomplete with less than the defined buffer period remaining before the deadline → P1
- Cross-session reconciliation finds a conditional task mentioned in a transcript with no corresponding structured field on the hire’s record → P2
- A new onboarding session type is deployed without a mandatory structured-field write step for conditional triggers → P3