ai-enabled-hr-talent-automation

Agent Architecture

Title: Agent Architecture

Version: 1.1

Owner: [TENANT_CONFIGURATION_REQUIRED — AI Governance Lead]

Status: Draft (target-state; most skills are unimplemented stubs — see notice)

Last reviewed: 2026-09-08

Next review: [TENANT_CONFIGURATION_REQUIRED]

Reviewers: AI Governance, Architecture, Security

Implementation reality notice (2026-09-08): there is no separate Supervisor Agent or Tool Gateway process. HrAutomation.Api controllers call HrAutomation.Application’s IWorkflowOrchestrator/ISkillRegistry in-process, which dispatch to ISkill implementations in HrAutomation.Agents. Of the skills this document describes, only CV extraction (parse_cv_skill) and the TAN create/approve skills (create_tan_skill, tan_approval_skill) have real logic; every other skill listed in agent-skill-catalog.md is a stub in HrAutomation.Agents/Stubs/ that always returns blocked/”not yet implemented.” HrAutomation.Mcp (the Tool Gateway/MCP servers this diagram shows) is an empty project scaffold — no MCP server exists to route to yet.

Purpose and scope

Details the agent orchestration pattern decided in ADR-002: a Supervisor Agent routing to Specialist Skills, all outside the deterministic Workflow Engine.

Architecture

flowchart TB
    subgraph Trigger
        E[Workflow event or API request]
    end
    E --> SUP[Supervisor Agent]
    SUP -->|route by task type| S1[CV Extraction Skill]
    SUP --> S2[Candidate Matching Skill]
    SUP --> S3[Interview Coordination Drafting Skill]
    SUP --> S4[Offer Drafting Skill]
    SUP --> S5[Document Verification Assist Skill]
    SUP --> S6[Employee Conversion Assist Skill]
    S1 --> TG[Tool Gateway - least privilege]
    S2 --> TG
    S3 --> TG
    S4 --> TG
    S5 --> TG
    S6 --> TG
    TG --> MCP[MCP Servers]
    S1 --> OV[Output Validator - JSON Schema]
    S2 --> OV
    S3 --> OV
    S4 --> OV
    S5 --> OV
    S6 --> OV
    OV --> API[HR Core API / Workflow Engine]
    API -->|approval-matrix check| DEC{Sensitive action?}
    DEC -- Yes --> HUMAN[Human approval required]
    DEC -- No --> COMMIT[Commit state change]

Supervisor Agent responsibilities

• Classify incoming task, select the allow-listed skill(s) for that task type only (no skill can be invoked outside its declared trigger context).
• Enforce per-skill token/timeout budgets (model-routing-and-cost-controls.md).
• Aggregate multi-skill results (e.g., matching may call extraction results) without exposing raw untrusted content across skill boundaries unnecessarily.
• Emit guardrail telemetry (confidence, tool calls made, approval state) per ADR-005.

Specialist skills

See agent-skill-catalog.md for the full catalog with inputs/outputs/tools/confidence thresholds.

Deterministic control boundary

The Supervisor and all skills run in the Agent Plane, architecturally separate from the Workflow Engine (see component-architecture.md). No skill holds a database credential capable of writing workflow-state tables directly; all state mutation happens through the HR Core API, which independently enforces the approval matrix regardless of what the agent recommends.

Failure handling

Skill timeouts, tool failures, or schema-validation failures result in a needs_review outcome routed to a human queue — never a silent fallback to an unvalidated default (see ai-guardrails-policy.md).

Change control

Version	Date	Author	Change
1.0	2026-09-07	Documentation package generation	Initial creation

This site is open source. Improve this page.