Description

arXiv:2604.17111 (April 2026) proposes HiveMind, an OS-inspired scheduling system for concurrent LLM agent workloads. It introduces primitives: admission control, provider-aware rate-limit tracking, and per-agent token budgets — directly applicable to Zeph's multi-agent PILOT routing layer.

Key Findings

• Admission control gate prevents overcommitting concurrent agents to a provider
• Provider-aware rate-limit tracking integrates with the provider registry (maps 1:1 to Zeph's [[llm.providers]] entries)
• Per-agent token budgets enforce cost ceilings per delegated subagent
• Scheduling policies (FIFO, priority, fair-share) selectable at runtime

Relevance to Zeph

Zeph's PILOT LinUCB bandit routing already tracks provider latency and cost but lacks:

1. Admission control (can saturate a provider with concurrent subagents)
2. Per-agent budget enforcement (subagents can exceed total turn budget)

HiveMind primitives could be layered on top of PILOT without replacing the bandit policy.

Research Actions

• Read full paper: https://arxiv.org/abs/2604.17111
• Prototype admission control gate in zeph-orchestration
• Assess whether per-agent token budgets can reuse cost_tracker already in zeph-core

Environment

• Paper: arXiv:2604.17111
• Area: zeph-orchestration, zeph-core