Source

arXiv:2603.13110 — AgentRM: An OS-Inspired Resource Manager for LLM Agent Systems (March 2026)

Technique

Applies classic OS scheduling theory to LLM agent systems in two components:

1. MLFQ Task Scheduler

• Multi-level feedback queue with priority lanes (interactive / batch / background)
• Rate-limit-aware admission control: tasks entering at high rate get demoted to avoid API throttle cascades
• Zombie reaping: detects and reaps stalled sub-agents to free scheduler slots
• Reduces P95 latency by 86% in multi-agent benchmarks

2. Three-Tier Context Lifecycle Manager (CLM)

• Tier 0 (Hot): active working memory — full fidelity, fast access
• Tier 1 (Warm): recent context — adaptive compaction applied
• Tier 2 (Cold): hibernated context — serialized to storage, recalled on demand
• Achieves 100% key-information retention vs. 65% for sliding-window baselines

Applicability to Zeph

High. Two independent components, both directly mappable:

• DagScheduler + zeph-scheduler would benefit from MLFQ priority lanes: /plan tasks as interactive, background compaction as batch, sweep timers as background. Zombie reaping maps to task_timeout_secs + explicit task cancellation.
• Zeph's context builder already has soft/hard compaction tiers; adding a hibernation tier (Tier 2) for very old sessions aligns with the CLM model. The deferred_apply_threshold mechanism is the admission gate for Tier 1→0 promotion.

Differs from #2320 (AdaptOrch) which focuses on topology selection (parallel/sequential/hierarchical DAG shapes); AgentRM focuses on resource management within a fixed topology.

Implementation sketch

1. Add priority field to TaskKind or PlanTask in zeph-orchestration
2. Implement admission control in DagScheduler::tick() based on current rate vs. configured limit
3. Add zombie detection: tasks exceeding task_timeout_secs get canceled and logged
4. Context hibernation: add hibernated tier below cold in memory tier config with SQLite serialization gate

P2 — research, evaluate MLFQ fit with current DAG scheduler before implementation.