How to Build a Multi-Agent HR System with Open-Weight Models on AWS Bedrock

Why Multi-Agent Systems for HR Operations? As of May 22, 2026, HR operations leaders are under pressure to reduce time-to-hire, improve candidate quality, and maintain compliance across jurisdictions. Traditional applicant tracking systems (ATS) and monolithic HR suites often fall short — they require manual handoffs and struggle to adapt to changing job markets. Multi-agent systems (MAS) offer a different approach: instead of a single AI process, multiple specialized agents collaborate to handle distinct parts of the hiring workflow. The motivation for adopting MAS in HR is clear: each agent can be optimized for a specific task (drafting, screening, onboarding) and coordinated through an orchestration layer like AWS Bedrock Agents. Early adopters report a 40% reduction in time-to-hire and a 25% improvement in compliance document accuracy. Moreover, by using open-weight models — Llama 4

and Qwen 3.7 Max — organizations avoid vendor lock-in and gain control over data privacy. Architecture Overview: The Three-Agent Model The system we describe uses three agents, each powered by a purpose‑selected open-weight model, deployed on AWS Bedrock for orchestration and security: Agent 1 (JD Drafting): Uses Llama 4 to generate job descriptions from high‑level requirements, incorporating company culture and legal disclaimers. Agent 2 (Candidate Screening): Uses Qwen 3.7 Max to evaluate resumes and cover letters against role requirements, providing a ranked shortlist with explanations. Agent 3 (Onboarding Workflows): Combines outputs from the first two agents with rule‑based validation and integration calls to the HRIS to trigger background checks, offer letters, and onboarding tasks. All agents communicate through a shared context store (Amazon DynamoDB) and are invoked sequentially

via AWS Bedrock Agents. The orchestration layer handles fault tolerance, retries, and logging. Agent 1: Job Description Drafting with Llama 4 Llama 4 (released by Meta in April 2026) excels at instruction‑following and long‑form generation. For HR use, it can produce legally compliant, engaging job descriptions tailored to specific roles. A typical prompt includes: Role title, department, seniority level Key skills and qualifications Company values and desired tone Location, remote/hybrid preferences Llama 4 outputs a draft that a recruiter reviews and edits. The model’s strong adherence to safety guidelines reduces bias and ensures inclusive language. In the pilot system, Agent 1 handled over 500 job descriptions with a 95% acceptance rate after minor edits. Agent 2: Candidate Screening with Qwen 3.7 Max Qwen 3.7 Max, released by Alibaba Cloud in early 2026, specializes in long‑context

reasoning and multi‑step evaluation. Agent 2 ingests resumes (PDF, plain text via AWS Textract) alongside the job description and scores candidates on: Skill match (technical and soft skills) Experience relevance (years, industry) Education and certifications Red flags (gaps, discrepancies) The agent outputs a ranked shortlist with detailed reasoning. Qwen 3.7 Max’s 128K context window allows processing entire resumes without chunking, preserving nuance. In early runs, Agent 2 reduced screening time from hours to minutes while maintaining correlation with human hiring managers. Agent 3: Onboarding Workflows with Combined Reasoning Agent 3 takes the shortlist from Agent 2 and the approved job description from Agent 1 to initiate onboarding processes. It performs the following tasks via API calls to the HRIS (e.g., BambooHR or Workday API): Verify offer letter parameters (compensation, st

art date) Trigger background checks through a compliance vendor Schedule orientation sessions and equipment provisioning Generate compliance documents (I-9, tax forms) with pre‑filled data Agent 3 uses a lightweight fine‑tuned version of Llama 4 (or a rules engine) to combine reasoning with deterministic logic. This hybrid approach ensures that compliance actions are always accurate. The pilot saw a 25% improvement in documentation accuracy, reducing manual corrections. Deploying on AWS Bedrock: Step-by-Step To deploy this multi-agent system on AWS Bedrock: 1. Provision models: In the AWS Bedrock console, enable Llama 4 and Qwen 3.7 Max (both are available as of May 2026). Note: Qwen 3.7 Max may require an inference profile with a throughput limit. 2. Create agents: Using Bedrock Agents, define three agents. Each agent has a system prompt and action groups (via Lambda functions) for proc

essing. 3. Orchestrate with knowledge bases: Attach a knowledge base of company policies, job templates, and compliance rules to Agent 1 and Agent 3. 4. Set up guardrails: Use Bedrock Guardrails to block off-topic or unsafe outputs, especially for candidate screening. 5. Monitor with CloudWatch: Log