Unified Harness Architecture Across “Where–When–What–With Whom” Extensions

Determine whether a single agent harness architecture can span the four axes of evolution—where the harness runs, when it acts, what it acts on, and with whom it coordinates—or whether specialized stacks are required.

Background

The paper surveys emerging directions that extend the harness along multiple axes: decoupled managed agents (where), proactive operation (when), multimodal/embodied action (what), and multi-agent collaboration paradigms (with whom). It is unclear if one architecture can accommodate all axes without fragmentation.

This question probes architectural generality versus specialization as agent capabilities and deployment contexts diversify.