Multi-Agent Harness for Long-Horizon Software Delivery

Keynote
Shanghai
3:50 p.m. - 4:30 p.m.
Venue B(Integrated Building, Room 434 Meeting Room)
  • Gongmin Liu Agent Architect

    Over 10 years of R&D and platform architecture experience in autonomous driving. He currently leads the design and development of the autonomous-driving Data Agent platform and the ZaoFu Coding Harness Platform, focusing on Agent Harness, Long-Horizon Agents, Multi-Agent collaboration, and AI-native software engineering. Author of the ZaoFu harness platform: https://github.com/uisee-ai/zaofu

    anonymous

Abstract

How Goals, Task Maps, dynamic Workflows, cross-provider Agents, controlled swarms, and independent verification turn scattered Agent sessions into an observable, recoverable, governable delivery system.

Details

Coding Agents already have strong code-generation ability, but what teams actually need is multiple Agents working toward the same goal across long-horizon tasks — continuously planning, implementing, verifying, redoing, and recovering.

Drawing on practice at ZaoFu, this talk covers how Goals, Task Maps, dynamic Workflows, cross-provider Agents, controlled swarms, and independent verification turn scattered Agent sessions into an observable, recoverable, governable software delivery system — and how a Run Manager and Autoresearch turn repeated failures into verified, cumulative Harness capability.