Agentic File System Abstraction for Context Engineering

• Source: https://arxiv.org/pdf/2512.05470
• arXiv: 2512.05470 (cs.SE)
• Author metadata (arXiv submission): Xiwei Xu
• Clipped: 2026-03-09 (SGT)

TL;DR

The paper argues that the bottleneck in GenAI systems has shifted from model tuning to context engineering. It proposes a Unix-inspired architecture (“everything is a file”) where memory, tools, external knowledge, and human inputs are mounted and governed through a unified file-system abstraction, then operationalized via a 3-part context pipeline: Constructor, Updater, Evaluator.

Core contributions

• Reframes context engineering as a software architecture problem, not just prompt design.
• Proposes an Agentic File System (AFS) abstraction to manage heterogeneous context sources under one namespace.
• Maps design to software engineering principles: abstraction, modularity, separation of concerns, traceability, composability.
• Defines a persistent context repository lifecycle with:
  • History (immutable log),
  • Memory (structured/indexed, mutable),
  • Scratchpad (temporary working context).
• Formalizes key GenAI constraints driving the architecture:
  • token-window limits,
  • model statelessness,
  • probabilistic/non-deterministic outputs.
• Introduces a closed-loop context pipeline:
  1. Context Constructor (select/score/compress context),
  2. Context Updater (inject/refresh context during runtime),
  3. Context Evaluator (validate outputs, update memory, trigger human review when needed).
• Implements the approach in the open-source AIGNE framework and demonstrates two exemplars:
  • memory-enabled agent,
  • MCP-based GitHub assistant.

Practical implications

• Pushes teams toward treating context as versioned, governed infrastructure (closer to DevOps/DataOps) instead of ad-hoc prompts.
• Improves auditability/provenance for enterprise AI systems.
• Supports human-in-the-loop co-work as a first-class design element.

Caveats / limits

• Mainly an architectural proposal + framework implementation; limited quantitative benchmark comparisons in the text.
• Evaluation is exemplar-driven (proof-of-feasibility) rather than broad empirical validation across many domains.
• Real-world performance depends on quality of retrieval, metadata discipline, and governance policy design.