[Paper] Bio-inspired Agentic Self-healing Framework for Resilient Distributed Computing Continuum Systems

Source: arXiv - 2601.00339v1

Overview

The paper presents ReCiSt, a bio‑inspired, agent‑based framework that brings self‑healing capabilities to Distributed Computing Continuum Systems (DCCS) – the sprawling ecosystems that span tiny IoT sensors, edge nodes, and massive cloud clusters. By mapping the four biological phases of wound repair (hemostasis, inflammation, proliferation, remodeling) onto computational layers, the authors show how autonomous agents powered by large language models (LLMs) can detect, diagnose, recover from, and learn about faults with only seconds of latency and modest CPU overhead.

Key Contributions

• Bio‑inspired architecture: Introduces a four‑layer model (Containment, Diagnosis, Meta‑Cognitive, Knowledge) that mirrors the body’s wound‑healing process.
• LLM‑driven agents: Leverages modern language models to parse heterogeneous logs, infer root causes, and generate remediation actions without handcrafted rules.
• End‑to‑end self‑healing loop: Demonstrates autonomous fault isolation, causal diagnosis, adaptive recovery, and long‑term knowledge consolidation in a single pipeline.
• Empirical evaluation on public fault datasets: Shows that ReCiSt can resolve incidents within tens of seconds while consuming ≤ 10 % of a CPU core per agent.
• Scalable micro‑agent orchestration: Quantifies how many lightweight agents are spawned to handle different fault scenarios, highlighting the framework’s ability to scale across the continuum.

Methodology

1. Mapping biology to software – The authors decompose the healing process into four computational layers:

   • Containment (hemostasis) isolates the faulty component.
   • Diagnosis (inflammation) gathers logs, metrics, and traces, then uses an LLM to hypothesize causes.
   • Meta‑Cognitive (proliferation) selects or synthesizes a recovery plan (e.g., restart a service, migrate a workload, re‑configure a network).
   • Knowledge (remodeling) stores the incident narrative and lessons learned for future reference.

2. Agent design – Each layer is implemented as a set of lightweight “micro‑agents” that communicate via a publish/subscribe bus. The agents are stateless except for the Knowledge layer, which maintains a vector‑store of incident embeddings for similarity search.

3. LLM integration – Prompts are crafted to turn raw log snippets into structured “symptom” objects, then into causal graphs. The same LLM can also generate remediation scripts (e.g., Kubernetes kubectl commands) that are validated before execution.

4. Evaluation pipeline – The framework is deployed on a testbed that mixes Raspberry‑Pi‑class edge nodes, a mid‑tier fog cluster, and a Kubernetes‑based cloud tier. Faults are injected from publicly available datasets (e.g., SMD, Yahoo! A3). Metrics captured include detection latency, CPU usage per agent, and the number of agents instantiated per incident.

Results & Findings

Even without a direct baseline (the authors note a lack of comparable self‑healing frameworks for DCCS), the numbers suggest that ReCiSt delivers fast, low‑impact recovery that scales with system heterogeneity.

Practical Implications

• Reduced on‑call fatigue – Developers can rely on autonomous agents to triage and fix many incidents, freeing human operators for higher‑level tasks.
• Edge‑to‑cloud resilience – Because the agents run on any node (from constrained IoT devices up to cloud VMs), the same self‑healing logic can be deployed across the entire continuum, eliminating the need for tier‑specific tooling.
• LLM‑as‑a‑service for ops – The work showcases a concrete, production‑grade use case for LLMs beyond chatbots: turning raw telemetry into actionable remediation.
• Knowledge‑driven incident management – The Knowledge layer creates a searchable “medical record” of system faults, enabling faster root‑cause analysis for recurring issues and supporting compliance/audit trails.
• Plug‑and‑play architecture – The micro‑agent model and pub/sub communication make it straightforward to integrate ReCiSt with existing observability stacks (Prometheus, OpenTelemetry, ELK) and orchestration platforms (Kubernetes, Nomad).

Limitations & Future Work

• Baseline scarcity – The authors could not compare against existing frameworks, making it hard to quantify relative gains.
• LLM dependency – Performance hinges on the quality and latency of the underlying language model; on‑prem LLMs may be required for privacy‑sensitive environments.
• Resource‑constrained nodes – While the reported CPU usage is modest, the memory footprint of LLM inference on ultra‑low‑power devices remains an open question.
• Security considerations – Automatically generated remediation scripts need robust sandboxing to avoid accidental destructive actions.

Future directions include:

1. Benchmarking against emerging self‑healing platforms.
2. Exploring model compression techniques for edge deployment.
3. Extending the Knowledge layer with reinforcement‑learning feedback loops.
4. Formal verification of agent‑generated actions.

Authors

• Alaa Saleh
• Praveen Kumar Donta
• Roberto Morabito
• Sasu Tarkoma
• Anders Lindgren
• Qiyang Zhang
• Schahram Dustdar
• Susanna Pirttikangas
• Lauri Lovén

Paper Information

• arXiv ID: 2601.00339v1
• Categories: cs.AI, cs.DC, cs.ET, cs.MA, cs.NE
• Published: January 1, 2026
• PDF: Download PDF