Persistent AI Agents in Academic Research: A Single-Investigator Implementation Case Study

Technical Summary: Persistent AI Agents in Academic Research

Problem Statement
Current evaluations of Large Language Models (LLMs) predominantly focus on isolated model performance, benchmark scores, or short conversational episodes. These approaches fail to capture the reality of how AI is increasingly embedded into continuous, real-world academic workflows. In such environments, agents operate with durable memory, access to local files and external tools, scheduled routines, delegated roles, and explicit safety protocols. There is a lack of quantitative, practice-based descriptions of these "persistent agentic research environments," particularly regarding their architecture, resource consumption, governance accumulation, and economic profiles over extended periods.

Methodology
The study presents a structured, self-observed implementation case study conducted over 115 days (January 31 – May 25, 2026) by a single academic physician-scientist.

• Unit of Analysis: The "human-agent environment," comprising the researcher, agent runtime, memory layer, tools, repositories, scheduled jobs, and governance rules.
• Data Sources: Recoverable session telemetry (de-duplicated), memory files, repository inventories, model-use logs, decision logs, and protocol updates.
• Measurement Framework: The authors introduced PARE-M (Persistent Agentic Research Environment Measurement) v0.1, a descriptive framework organizing metrics into five domains:
  1. Utilization: Active-day fraction, de-duplicated record counts, message roles, tool calls, and active system time (estimated via capped-gap algorithms).
  2. Outputs: Output-proxy events (memory entries indicating completed artifacts) and artifact-surface breadth.
  3. Resources: Token telemetry (input, output, cache-read, cache-write) and runtime infrastructure specifications.
  4. Reproducibility: Data schemas, parsing rules, and de-duplication logic.
  5. Governance: Rates of failure, verification, correction, and protocol-proxy events.
• Constraints: The study is not a controlled productivity trial; it lacks a control group, baseline productivity period, or independent coding for governance events. Claims are descriptive rather than causal.

Key Results

• System Architecture & Utilization: The environment accumulated 502 memory-related files, 17 configured agent directories, and 57 skill files. Over 96 active days (83.5% of the observation window), the system generated 75,671 de-duplicated records, including 8,059 user messages, 23,710 assistant messages, and 2,385 tool-call events. Active system time was estimated at 579.7 hours (using a 30-minute gap cap).
• Outputs & Governance: The system produced 482 output-proxy events (5.02 per active day) across 10 distinct artifact categories (manuscripts, teaching materials, software, operations, etc.). Crucially, the governance layer was active and integrated, recording 889 failure, verification, correction, or protocol-proxy events (9.26 per active day).
• Resource Profile (Token Telemetry): Analysis of a strict May 1–25 subset (627 model-completed events) revealed a cache-dominant workflow. Of the 73.95 million recorded tokens, 61.28 million (82.9%) were cache-read tokens. Input tokens accounted for ~10.7 million, and output tokens for ~0.75 million.
• Economic Implications: Preliminary direct spend was approximately US$1,961. However, the high cache-read ratio suggests that "cost per token" is a poor proxy for value in persistent workflows, as the system relies heavily on reused context rather than fresh inference generation.

Key Contributions

1. Empirical System Description: Provides the first quantitative, longitudinal description of a persistent AI agent operating within a real academic research workflow, moving beyond isolated chatbot interactions.
2. PARE-M Framework: Introduces a measurement framework (PARE-M) specifically designed to evaluate persistent agentic environments, emphasizing artifact-level denominators, reproducibility rules, and governance event tracking.
3. Cache-Dominance Evidence: Demonstrates that in persistent workflows, the majority of token usage consists of cache reads (82.9%), indicating that the economic unit of value shifts from "cost per token" to "cost per completed artifact" or "cost per verified workflow."
4. Governance Integration: Shows that safety and governance protocols (e.g., citation verification, deployment checks) accumulate within the memory layer as an operational part of the environment, rather than existing as separate policy appendices.

Significance and Claims
The paper modestly claims that persistent agentic infrastructure expands the capacity and scope of academic work, rather than proving a direct reduction in human labor input. The primary finding is that as the environment accumulates memory, tools, and procedures, the scope of delegated work expands.

The authors argue that episodic benchmarks (e.g., HELM, SWE-bench) miss the phenomenon of lived deployment. Future evaluations should move away from raw token volume and toward:

• Artifact-level efficiency metrics.
• Cost-per-artifact estimates.
• Independent coding of governance events.
• Reproducible parsing rules for system telemetry.

The study concludes that while the system did not demonstrate labor substitution in this single-investigator case, it established a durable work system spanning research, teaching, software, and operations, necessitating new standards for measurement, cost accounting, and safety governance in AI-assisted research.