Multi-Agent AI System Automates Special Education Planning

Multi-Agent AI System Automates Special Education Planning

A new AI application is transforming how schools create transition plans for students with disabilities, cutting a traditionally labor-intensive process down to minutes while maintaining strict legal compliance.

Trinity, developed by University Startups in partnership with AWS consulting firm g/d/n/a, uses a six-agent architecture built on Amazon Bedrock to guide students through creating Individualized Education Program (IEP) transition plans. The system is now deployed across more than a dozen U.S. states and expanding internationally to Saudi Arabia and Kuwait.

Why it matters

Transition planning under the Individuals with Disabilities Education Act (IDEA) requires coordinating career exploration, postsecondary education options, and independent living goals—a process that typically demands significant educator time and specialized knowledge. By automating the technical aspects while keeping students at the center of the conversation, Trinity addresses both compliance requirements and the practical reality of limited special education resources. The architecture also demonstrates how breaking complex AI workflows into specialized agents can improve accuracy and maintainability at scale.

From monolithic to multi-agent

The breakthrough came when the development team abandoned a single-model approach in favor of specialized agents, each handling a distinct domain. An orchestrator agent coordinates five specialist agents: one for college recommendations drawing from 30,000 institutions, another for employment pathways across 1,000 occupations, a third for vocational training programs from a 5,000-record database, plus agents for community engagement and independent living skills.

This separation solved a critical problem in the early prototypes. When student intake, career exploration, compliance checking, and plan writing all competed in one prompt, failures in any domain could corrupt entire sessions or produce hallucinations. The multi-agent design isolates these functions while maintaining consistency through a shared student profile stored in Amazon DynamoDB.

The system operates in two phases. First, students complete a conversational intake covering their interests, strengths, and goals. The orchestrator routes to the appropriate primary agent, which returns five to ten filtered recommendations. In the second phase, after students confirm their selections, all relevant agents contribute sections that the orchestrator consolidates into a structured, IDEA-aligned plan exportable as a formatted PDF within five to ten seconds.

Hybrid retrieval and compliance controls

Trinity's three Amazon Bedrock Knowledge Bases use a weighted scoring model that balances semantic similarity (30 percent), keyword matching (25 percent), location preference (20 percent), and program attributes including disability support availability, cost, and education level (25 percent). Results below a 25 percent relevance threshold are filtered out entirely.

Compliance requirements shaped the architecture from the start. Field-level encryption through AWS Key Management Service protects personally identifiable information including names, disability types, and accommodations, with keys rotating annually. Role-based access control enforces data boundaries: district coordinators see only their district's data, school staff view their school's information, and students access only their own sessions.

Authentication integrates directly with Canvas learning management system using LTI 1.3, eliminating separate credentials while AWS manages the flow through a custom Lambda-backed OpenID Connect provider.

Field results

Early adopters report measurable efficiency gains. Educators can now complete student survey, career discovery, and compliance documentation in under ten minutes per student—a process that previously required distributed resources and manual coordination across multiple sessions.

Accessibility features including Amazon Polly for text-to-speech and Amazon Transcribe for voice input ensure students with varying abilities can engage with the system independently.

These details were first reported by AWS in a blog post co-authored with University Startups and g/d/n/a.