North-East AI Agents Day 2026, 2026

Kourai Khryseai: Transparent Human-on-the-Loop Multi-Agent Software Development

A. Barea

Abstract

Multi-agent coding systems are increasingly capable, but their coordination decisions are often opaque to the supervisor. Kourai Khryseai treats multi-agent software development as an interpretability problem and organizes the system around three pillars: Monitor, Communicate, and Control. A Hephaestus orchestrator routes requests to specialist agents for planning, coding, testing, style review, and commit synthesis. Every step is observable through OpenTelemetry GenAI spans and a streamed Forge transcript (Monitor); agents pause to request clarification when requirements are ambiguous (Communicate); and a bounded Kallos⇅Techne repair loop plus graceful degradation make recovery deliberate rather than an edge case (Control). The backend combines independent A2A-connected services, MCP-based tool use, shared local SQLite state, and end-to-end tracing through OpenTelemetry, Jaeger, and Prometheus. The same orchestration layer powers a CLI, desktop GUI, and visual-novel-style interface, enabling studies of interface and embodiment without changing backend logic. Claim-driven tests anchor each MCC property to a mechanism in code: transparency is treated as a systems property, not just a UX choice.

Lab authors

  • AJ Barea

    RIT PhD Researcher · Sphragis · Pharos · Phalanx-FL