Recoleta
GitHub
Source note

GENESIS: Harnessing AI Agents for Autonomous 6G RAN Synthesis, Research, and Testing

Authors
Tamerlan Aghayev, Maxime Elkael, Michele Polese, Minh Dat Nguyen, Gabriele Gemmi, Andrea Lacava, et al.
Source
arXiv
Published
Collection
Software Intelligence
Multi Agent Software EngineeringCode GenerationRan EngineeringSpec To CodeAgentic Testing
Open arXivMarkdown source

Summary

Genesis is a multi-agent system for turning 3GPP/O-RAN intents into RAN code, tests, patches, and experiments that run on simulators, emulators, and real radios. The paper claims 100% success on two RAN feature-implementation case studies, while Claude Code with Opus 4.7 produced no working implementation.

Problem

  • RAN R&D is slow because feature synthesis, conformance testing, hardening, optimization, discovery, and security work each take months of engineering time per iteration.
  • In one cited 5G stack analysis, a substantial feature took 74 days on average to reach the stable branch, with a 207-day 90th percentile.
  • General LLM coding agents fail on RAN tasks because a small API error or spec misread can break interoperability with standard devices, radios, and OTA testbeds.

Approach

  • Genesis routes an intent, such as a spec clause, telemetry anomaly, or research hypothesis, into one of 6 pipelines: Synthesize, Test, Harden, Optimize, Discover, or Secure.
  • The core mechanism is simple: agents decide what to do, deterministic skills run actions such as build, deploy, configure, and experiment, and hooks log events, enforce safety gates, and record provenance.
  • Synapse stores curated 3GPP/O-RAN specs, research papers, reference code, lab inventory, code diffs, logs, traces, and experiment outputs, so later runs can reuse validated artifacts.
  • Validation moves across 3 tiers: RFSIM simulation, Colosseum or hardware-in-the-loop emulation, and OTA deployment on production-grade testbeds such as X5G and Arena.
  • The prototype uses about 23 parameterized skills and can run commercial and local models, including Claude Opus 4.7, Claude Sonnet 4.6, gpt-oss, Llama 4, Phi, and Gemma.

Results

  • Across multiple statistically independent runs, Genesis achieved 100% success implementing RRC.ConnMean KPM from 3GPP TS 28.552 and Conditional Handover with a closed-loop E2SM-RC xApp.
  • The baseline, Claude Code with Opus 4.7, produced 0 working implementations on every attempt for the first 2 case studies.
  • The evaluation covers 3 end-to-end case studies: RRC.ConnMean KPM synthesis and OTA propagation, CHO synthesis/testing/hardening, and autonomous RAN scheduler discovery using the ALLSTaR loop.
  • The system covers the full 6-step RAN R&D cycle through 6 pipelines and validates generated changes on 3 infrastructure tiers before feeding results back into Synapse.
  • The paper reports profiling across stages and says 2 of the 6 Synthesize stages, feature implementation and test execution, dominate both cost and wall-clock time.

Link