research(arch): empirical agent harness study — audit Zeph against 5 recurring design dimensions (arXiv:2604.18071)

[bug-ops]

Description

Paper: "Architectural Design Decisions in AI Agent Harnesses" (arXiv:2604.18071, April 2026). Empirical analysis of 70 open-source agent systems, identifies five recurring design dimensions.

Five Design Dimensions

1. Subagent architecture — how agents spawn, isolate, and coordinate children
2. Context management — file-persistent, hybrid, and hierarchical strategies dominant
3. Tool systems — registry-oriented dominant; MCP- and plugin-oriented emerging
4. Safety mechanisms — "intermediate isolation is common but high-assurance audit is rare"
5. Orchestration — deeper coordination pairs with more explicit context services

Relevance to Zeph

The paper finds that "stronger execution environments correlate with more structured governance" and "formalized tool-registration boundaries align with broader ecosystem ambitions."

Gap identified: "High-assurance audit is rare" — Zeph has tool audit logging but it is not verified as a systematic safety invariant. The new hooks.rs (+328 lines) is the right place to enforce this.

Acceptance Criteria

• Evaluate Zeph against each of the five dimensions
• Identify where Zeph sits in the taxonomy (multi-agent orchestrator vs enterprise system)
• File issues for any structural gaps found, particularly around safety audit coverage

References

• Paper: https://arxiv.org/abs/2604.18071
• Related: arXiv:2604.24657 (AgentWard, research(security): AgentWard lifecycle security — map five-stage defense model to zeph-subagent hooks.rs (arXiv:2604.24657) #4012)