Research

Ongoing investigations into AI agent reliability, hardware optimization, and system design patterns.

Agent Self-Harness Loop Detection

Based on arxiv:2606.09498 — agents can detect their own failure patterns and correct structurally. Implemented as a hard rule: ≥3 consecutive turns same action type forces pivot. Prevents infinite loops without external monitoring.

SOUL.md + AGENTS.md · Self-Harness paper reference

Temporal Correlation ≠ Causation

Same-day co-occurrence doesn't prove causation. Line up timestamps of X and Y before claiming X→Y. Critical for operational debugging where multiple events happen during outages.

claude-code memory · Config-drift incident lesson

Config Drift Prevention

When multiple agents modify shared config, drift accumulates silently. Solution: atomic writes, pre-commit validation, and explicit attribution in code comments. Never fake-attribute decisions to humans.

scripts/atomic-write.sh · SOUL.md Non-Negotiable

Safe Restart Protocol

Direct systemctl stop on inference engines = self-deletion. Service restarts must go through safe-restart with queuing, idle-wait, and user notification. Six-hour outage resulted from violating this rule.

MEMORY.md · SOUL.md Non-Negotiable

GPU Power Limit Experiment

Explored RTX 3090 power limits, persistence mode, PCIe register probing. Found that manual power limit changes under load cause instability. Lesson: never touch GPU settings without explicit authorization and rollback plan.

nvidia-smi read-only queries only · Hardware intervention ban