UACP-HMA

โปรโตคอลสื่อสารกลางสำหรับระบบเอเจนต์ต่างชนิด

Universal AI Communication Protocol for Heterogeneous Multi-Agent Systems — a transport-agnostic message contract that lets models from different vendors and runtimes cooperate on one task graph.

โปรโตคอลสื่อสารกลางที่ทำให้โมเดล AI ต่างค่ายและต่างรันไทม์ทำงานร่วมกันบนกราฟงานเดียวกันได้

View ResearchDemoArchitecturePaperDatasetCodeAPI

Problem

Every agent framework invents its own message envelope, so combining a cloud reasoning model with a local Thai LLM and a vision service requires bespoke glue code that breaks on every upgrade.

Research Question

Can a single capability-negotiated message contract preserve task fidelity across heterogeneous models without central vendor lock-in?

Objectives

Define a minimal, versioned envelope for task, context, evidence and verdict messages
Support capability discovery so new agents can join at runtime
Preserve provenance and verification chains across model handoffs
Benchmark protocol overhead against direct API orchestration

Methodology

01Protocol specification and formal message schema
02Reference implementation with adapter layer per provider
03Simulation harness with synthetic multi-step task graphs
04Ablation on verification and retry strategies

Architecture

Client / UserWeb, API

Submits a task goal

API GatewayFastAPI

Authentication, quota, audit log

AI OrchestratorUACP-HMA core

Task graph planning and routing

Agent LayerHeterogeneous agents

Research, coding, vision, Thai language

Model LayerCloud + local LLM

Provider adapters, no direct UI coupling

Knowledge LayerVector + relational

Shared evidence store

ResultVerified output

Provenance chain attached

Results & Benchmark

Protocol overhead: measured in simulation
Agents interoperated: 7 adapter types

Limitations

· Results are from simulation harnesses, not production traffic
· Adapter coverage is limited to OpenAI-compatible and local runtimes so far