Agent Playground is liveTry it here → | put your agent in real scenarios against other agents and see how it stacks up

At a Glance

Agent teams operating on a live virtual copy of the network can automatically generate and run intent-aware tests that catch nearly all critical change errors far faster than manual lab work.

Key Findings

Aether uses multiple AI agents plus a network “twin” (a live, queryable copy of network state) to translate human change intent into focused test plans and run them automatically. Across synthetic benchmarks and two real ISP incidents it found nearly all critical problems that standard CI/CD checks miss, while keeping turnaround time low. The system delivers high test coverage and consistency, though it tends to be conservative and produces some false positives that need human review. This aligns with the ReAct Pattern (Reason + Act) for iterative reasoning.

Data Highlights

194% error detection across synthetic benchmarks and 100% detection on two real ISP incidents
2Automatic test coverage typically between 92%–96%; per-run averages in the ISP cases were ~91.7% and 95.6%
3Average end-to-end validation run ≈ 6 minutes, with ~55% of time spent running verification tools and ~45% on agent reasoning and orchestration

Why It Matters

Network engineers, SREs, and technical leads running large or critical networks — especially ISPs and enterprises with frequent config changes — can use this approach to reduce outages and speed safe rollouts. Teams building NetOps pipelines or CI/CD for network changes can plug in agent-driven validation to add intent-aware checks that standard automation misses. As seen in Multi-Agent Crisis Management deployments, such patterns support resilient, intent-aware validation workflows.
Not sure where to start?Get personalized recommendations
Learn More

Key Figures

Figure 1: Network Change Validation process.
Fig 1: Figure 1: Network Change Validation process.
Figure 2: Aether workflow.
Fig 2: Figure 2: Aether workflow.
Figure 3: Agents logical architecture.
Fig 3: Figure 3: Agents logical architecture.
Figure 4: NDM Knowledge Graph structure.
Fig 4: Figure 4: NDM Knowledge Graph structure.

Ready to evaluate your AI agents?

Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.

Learn More

Yes, But...

Accuracy depends on a high-quality, up-to-date network knowledge graph and correct schema ingestion; poor inputs degrade results. The system is conservative: detection rates are high but precision is moderate (leading to false positives that require human triage). CI/CD and operational integration were demonstrated but not exhaustively tested at production scale; protocol-specific skills and tuning remain important. This is also discussed in the Red Teaming Pattern.

The Details

Aether combines multiple AI agents with a Network Digital Twin (a time-indexed graph of the live network) and a suite of specialized verification tools. Agents read human change intent (ticket, pull request), query the knowledge graph for relevant state, generate focused test plans, and execute those tests inside the twin so production is never touched. The architecture separates reasoning (agents) from correctness checks (verification tools), letting the agents compose precise, intent-aware verifications instead of running broad, expensive full-matrix tests. Evaluation used 8 benchmark scenarios and two real ISP incident recreations on a lab replica (25 routers, hundreds of VRFs and ACLs). Results show strong practical value: near-total detection of critical protocol and policy failures (94% in benchmarks, 100% in the two production incidents), high coverage (92–96%), and fast runs (~6 minutes), making the flow suitable for CI/CD gates. Key limitations include moderate precision (false positives), fragility when the knowledge graph schema is complex or incomplete, and the need to inject protocol-specific expertise as agent skills to handle subtle networking logic reliably. The system is designed Human-in-the-Loop so operators retain final approval while routine validation is automated. This architecture aligns with the Evaluation-Driven Development (EDDOps) approach and benefits from the Orchestrator-Worker Pattern to separate orchestration from execution.
Test your agentsValidate against real scenarios
Learn More
Credibility Assessment:

Authors affiliated with established industry firms (Cisco, Swisscom) and one author with h-index 11, but venue is arXiv and no top-tier conference or very high h-indices—solid but not top-tier.