Christopher Aytona

Christopher Aytona · 2026 · Zenodo / SSRN (preprint)

Large language model agents deployed in persistent, multi-session environments face compounding challenges: memory degrades without governance, coordination between concurrent agents lacks formal guarantees, skill acquisition proceeds without quality control, and self-improvement operates without safety bounds. We present SHARD (Self-Healing Agent with Resilient Delegation), a harness-agnostic application architecture that composes four independent subsystems—memory governance with staleness detection, intent-based coordination protocol, quality-gated skill lifecycle with trust tiers, and safety-constrained self-improvement—into an integrated infrastructure where emergent behaviors arise from their interaction.

Christopher Aytona · 2026 · In Progress

Persistent LLM agent systems consume unbounded resources as context windows fill with tool schemas, memory segments, and session history. This work reframes resource constraints as architectural primitives rather than limitations, proposing three mechanisms — lazy MCP loading, tiered context injection, and zero-token routing — that reduce per-session overhead by 60–75% while preserving capability. Builds on the SHARD framework to make persistent agent architectures economically sustainable at scale.