Applied Patterns

1 · New-hire onboarding orchestrator

Goal: turn an accepted Ashby offer into a complete, ready-to-approve Workday hire + provisioning plan, with HRBP / Payroll signoff before commit.

Trigger

• Ashby webhook: offer accepted (or status → "Onboarding")

Inputs (reads, idempotent)

• ashby_get_application(application_id) — candidate, offer terms, start date
• ashby_get_offer(offer_id) — salary, bonus, equity, sign-on, location
• workday_get_position(position_id) — confirm the slot is still open
• workday_get_supervisory_org(sup_org_id) — manager, location defaults
• policy_get(country=candidate.country, topic="new_hire") — local statutory requirements
• payroll_get_calendar(country) — confirm start date is safely outside any lock window
• workday_check_existing_worker(email, ssn_hash) — rehire detection

Agent reasoning (Sonnet, tools attached)

System prompt: "Draft a complete new-hire package. Validate the start date against payroll calendar. If the candidate is a rehire, surface that prominently. Propose, do not commit."

Proposed actions (write-MCP, non-committing)

• propose_workday_hire — full Hire BP payload as an EIB-ready row
• propose_okta_provision — group memberships per role + location
• propose_equipment_order — laptop SKU per country
• draft_welcome_email — to candidate + buddy + manager
• propose_calendar_invites — day-1 orientation, week-1 1:1s
• propose_benefits_enrollment_window — per-country rules

HITL gates

• HRBP approves the Workday hire payload
• Payroll Ops approves if start date is within 5 business days of cutoff
• IT auto-approves Okta provisioning (low-risk, reversible)

Commit

Approved proposals execute via n8n: EIB upload runs Hire BP, Okta API provisions, calendar invites send. Idempotency keys ensure retry safety.

Audit log entry shape

{
  "run_id": "onb-2026-05-12-001",
  "trigger": {"source": "ashby", "event": "offer_accepted", "id": "ASH-9921"},
  "model": "claude-sonnet-4-5",
  "system_prompt_hash": "sha256:1f3a...",
  "tools_invoked": [
    {"name": "ashby_get_offer", "ts": "...", "result_hash": "..."},
    {"name": "workday_check_existing_worker", "ts": "...", "result": "no_match"}
  ],
  "proposals": [
    {"id": "P-101", "kind": "workday_hire", "idempotency_key": "hire:ASH-9921:2026-06-01"},
    {"id": "P-102", "kind": "okta_provision", "idempotency_key": "..."}
  ],
  "approvals": [
    {"proposal_id": "P-101", "approver": "hrbp:janet@", "ts": "...", "decision": "approve"}
  ],
  "commits": [
    {"proposal_id": "P-101", "executed_at": "...", "result": "eib_run_id:EIB-44919"}
  ],
  "pii_classification": "Confidential"
}

2 · Leaver checklist agent

Goal: when a termination is initiated, produce an end-to-end offboarding plan that catches every dependency — access, equity, final pay, knowledge transfer.

Trigger

• HRBP initiates Terminate BP in Workday → outbound webhook

Inputs

• workday_get_worker(employee_id) — current state
• workday_get_compensation(employee_id) — base, bonus, equity
• policy_get_severance(country, role_band, tenure)
• workday_list_pending_actions(employee_id) — open BPs, future-dated changes
• jira_list_open_tickets(reporter=employee_id) — knowledge transfer scope
• okta_list_groups(user) — access to revoke
• payroll_get_calendar(country)
• equity_get_grants(employee_id) — vesting, treatment per plan

Proposed actions

• propose_workday_termination — effective date, reason, last day worked
• propose_final_pay_calc — PTO payout, prorated bonus, severance
• propose_equity_treatment — accelerate / forfeit / convert per plan
• propose_okta_offboarding — staged revocation (last day vs. immediate)
• draft_offboarding_communication — to manager, to team, to candidate
• create_knowledge_transfer_jira
• propose_jurisdictional_notices — e.g. WARN in US, redundancy notice in UK

HITL gates (two-key for payroll-affecting steps)

• HRBP approves termination + comms
• Payroll Ops approves final-pay calc (separate approver)
• Legal approves jurisdictional notices when present

3 · Employee Q&A agent over policies

Goal: answer "how does X work for me?" questions from employees in Slack — accurately, with sources, with jurisdiction-awareness.

Trigger

• Slack mention in #ask-people or DM to People bot

Inputs

• workday_get_worker(slack_user_id → employee_id) — country, location, role, tenure
• policy_retrieve(question, country=worker.country) — RAG over policy library, jurisdiction-filtered
• workday_get_balances(employee_id) — PTO, sick balance if relevant

Agent reasoning

System prompt: "Answer the employee's question using only retrieved policy chunks for their country. Cite each fact with the policy doc reference. If the question requires a Legal interpretation, say so and route to HRKX."

Output

• Structured response: {answer, citations[], confidence, route_to_human}
• If route_to_human or confidence < threshold → HITL queue
• Else → post to Slack with citations

Eval surface (heavy)

• Accuracy on labeled policy Q&A set per jurisdiction
• Citation correctness (every claimed fact maps to a real chunk)
• Jurisdictional cross-talk (UK question never answered with US policy)
• PII leakage tests (employee A can't get answers referencing employee B's data)

See 07-evaluation-quality for eval design.

4 · Survey synthesis agent (Qualtrics)

Goal: at survey close, produce a leadership-ready synthesis of free-text responses — themes, sentiment, risk flags.

Trigger

• Scheduled run after Qualtrics survey closes

Pipeline (multi-agent appropriate here)

1. Clusterer (Sonnet) — embed responses, cluster by theme. Returns labeled clusters.
2. Sentiment scorer (Haiku, per-response) — high-volume, simple. Returns sentiment + intensity.
3. Risk flagger (Sonnet) — scans for red-flag content: harassment, retaliation, regulatory issues, leadership concerns. Routes flagged content to HRBP immediately, before synthesis.
4. Narrative writer (Opus) — given clusters + sentiment + cross-cycle context, drafts a leadership summary.

HITL gate

• People Analytics lead reviews and approves before distribution
• Risk-flagged content is routed immediately to a designated HRBP, never goes through the synthesis path

5 · Promotion-cycle analytics agent

Goal: support promotion committees by pre-scoring packets against the rubric, surfacing edge cases, and producing a calibration view across the org.

Trigger

• Manual kickoff by People Analytics when cycle opens

Inputs

• Promotion packets (from a Google Drive / Notion folder)
• Role rubric for target level (from policy library — prompt-cached)
• Performance review history (from Workday Talent module)
• Tenure-in-role, scope, comp band data (Workday)

Agent reasoning

For each candidate: score each rubric dimension (qualitative + cited evidence), flag inconsistencies (e.g. self-assessment 5/5 but manager 3/5), produce a recommendation tier (clear yes / clear no / committee discussion).

HITL gate

• The agent's output is advisory. Committees decide.
• Edge cases are surfaced for explicit discussion ("this candidate's rubric scores diverge sharply from manager assessment").

6 · Recruiter pipeline summary agent

Goal: produce a weekly hiring-health snapshot per recruiter / function from Ashby pipeline data.

Trigger

• Weekly cron, Friday mornings

Inputs

• ashby_list_active_jobs()
• ashby_pipeline_metrics(job_id, range) — applications, screens, onsites, offers, stage durations
• ashby_list_offers_pending()
• Prior-week snapshot for delta computation

Output

• Markdown summary per recruiter: open reqs, pipeline funnel, stage-conversion rates, offers out, deltas vs prior week, flagged stale candidates (no movement in >14d), at-risk reqs
• Posted to Slack channel per function; archived to a Notion page

HITL gate

• None — low-risk read-only summary. Sampled review (5%) for quality.

Reusable scaffolds — the templates HRKX inherits

"Multiply the People team's capability" is in the JD. Concretely you deliver, for each pattern above:

• Runbook — what the agent does, what it doesn't, what to do when it fails, who owns it
• n8n template — exportable JSON the team can fork to spin up variants (e.g., a contractor-onboarding variant of the new-hire pattern)
• Eval set — the labeled examples used to validate the agent, owned by HRKX going forward
• Risk register entry — tier, owner, halt criteria, last review date
• Replay protocol — given a run_id, how to reproduce and inspect