NE Agents Day 2026 — Kourai Khryseai¶
Title: Kourai Khryseai: Transparent Human-on-the-Loop Multi-Agent Software Development
Author: AJ Barea, Rochester Institute of Technology (ajb6289@rit.edu)
Advisor: Dr. Leon Reznik
Venue: North-East AI Agents Day 2026
Date: Friday, May 8, 2026
Location: Jane Street, New York, NY
Downloads:
- kourai-khryseai-poster.pdf — conference poster (48 × 36 in, print quality)
- kourai-khryseai-extended-abstract.pdf — accepted extended abstract

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.
Keywords: AI agents, multi-agent systems, software engineering, human-on-the-loop, interpretability, observability
Try it¶
The full system is on GitHub at
ajbarea/kourai-khryseai.
Clone, set an ANTHROPIC_API_KEY in .env, and:
make setup # uv workspace install
make up # build + start all 10 agents + vn-bridge + observability
make cli # interactive REPL
The three pillars (Monitor / Communicate / Control) map to verified
code locations documented in
Architecture → The Three Pillars.
For agentic-coding tools (Claude Code, Cursor, Codex, Aider), the root
AGENTS.md
is the canonical entry point; make help enumerates every developer
target.
Cite this work¶
@inproceedings{barea2026kourai,
author = {Arnaldo Barea},
title = {Kourai Khryseai: Transparent Human-on-the-Loop Multi-Agent Software Development},
booktitle = {North-East AI Agents Day 2026},
address = {New York, NY, USA},
year = {2026},
month = {5},
url = {https://ajbarea.github.io/kourai-khryseai/research/ne-agents-day-2026/},
}
The repo's
CITATION.cff
is the machine-readable source — GitHub renders a one-click "Cite this
repository" widget on the repo sidebar that exports BibTeX, APA,
EndNote, and Zenodo formats.