Agent architectureⓘ

Boundary rule: something is its own agent only when it owns a distinct decision, an isolated failure mode, or different oversight. Everything else is a skill.

Human checkpoints (HITL)Built

Consequential advance / reject decisions are recorded human actions (not automated); the fraud gate blocks advancing a flagged candidate. Every decision is in the audit log.

Applied→Fraud gate→Matching (deterministic)→Engagement / screening→HITL advance/reject

Agents (3)

TA SupervisorBuilt

Single accountable controller — owns workflow state, stage sequencing, the fraud-gate-before-matching rule, HITL checkpoints and the audit trail.

ta_orchestrationta_skill_matchingta_sourcing

Recruiters (Spike Chat) · no candidate channel

Authenticity & FraudBuilt

Independent, upstream of matching. Flags AI-generated CVs, fabricated history, duplicates/rings, identity anomalies. Flags for human review — never rejects, never scores fit.

ta_fraud_detection

Supervisor + recruiters · no candidate channel

Engagement — AriaBuilt

Conversational interface. Candidate Q&A, job recommendations, screening administration and scheduling. Assessment/grading live here as skills.

ta_candidate_experienceta_screeningta_scheduling

Candidates (web chat + WhatsApp)

Deterministic models & operations

The fit score stays an auditable model; agents orchestrate and explain it.

Fit-score modelBuilt

Deterministic, weighted skill-overlap — explainable, not an LLM judgment. Per-line calculation shown on each requisition.

schedule_interview / reschedule_interviewBuilt

Contended writes — no double-booking, optimistic version lock.

invite_screeningBuilt

Creates the one-way screening + question set.

Guardrail layer (not agents)

Compliance / oversight sits as a layer, not an autonomous agent.

ExplainabilityBuilt

Fit-score derivation shown per candidate; agents must cite evidence.

Human-in-the-loop gatesBuilt

Advance/Reject is a recorded human decision; fraud gate blocks automated advance of flagged candidates. See Governance.

Audit logBuilt

Every decision written immutably with actor + rationale + stage transition. See Governance.

Bias monitoringPartial

Four-fifths adverse-impact read-out by source/country; full protected-attribute monitoring needs consented data.

EU AI Act (high-risk hiring)Partial

Oversight, logging, explainability in place; candidate transparency partial. See Governance.

Deferred splits: Sourcing & Parsing becomes its own agent when external public-source retrieval, inbound-CV parsing and a skills taxonomy (ESCO/O*NET) land — until then the isolation argument doesn't apply and it stays the Supervisor's ta_sourcing skill.