Finding

Paper: "AI Runtime Infrastructure"
arXiv: https://arxiv.org/abs/2603.00495

Core Idea

Proposes an execution-time layer that sits between the orchestrator and agent execution to:

• Observe and intervene in agent behavior (not just pre/post hooks)
• Adaptive memory management (evict, compress, or retrieve based on live context pressure)
• Failure detection and recovery (retry strategies, alternative plan branches)
• Policy enforcement at runtime (not just at prompt time)

Key insight: separating the runtime infrastructure concern from agent logic enables agent-agnostic reliability improvements.

Applicability to Zeph

High (4/5). Zeph's architecture already has elements of this:

• agent loop as orchestrator
• for memory management
• Anomaly detection in
• Context assembly in

Gap: these are statically wired, not a composable runtime layer. The paper's framing suggests refactoring toward a pluggable runtime layer that intercepts between orchestrator decisions and tool/LLM calls — enabling:

1. Adaptive context pressure response (not just reactive compaction thresholds)
2. Per-tool failure recovery policies (retry, fallback, skip)
3. Memory eviction guided by live context budget, not just turn count

Implementation Sketch

Add a trait in that wraps tool dispatch and LLM calls, callable from the agent loop with access to current . Initial implementation: no-op passthrough. Future: plug in adaptive policies.