GitHub for RAN Intelligence: Inside GSMA's AI Agents for RAN Challenge and the Push Toward Verifiable Autonomous Networks

GitHub for RAN Intelligence: Inside GSMA’s AI Agents for RAN Challenge and the Push Toward Verifiable Autonomous Networks

Bottom line up front: On 10 November, at an event in Doha, the GSMA’s Open Telco AI initiative—working with ETSI, IEEE, ITU and Huawei—announced two competitions: TelcoRetrieve and the AI Agents for RAN Challenge. The second is the one that matters for anyone building or operating radio networks. It asks participants to build AI agents that reason, use tools and solve real radio engineering tasks inside Giulia, an open-source radio network simulator developed by Universitat Politècnica de València. The prize pool is €30,000. A webinar on 18 November will explain the mechanics.

Our view: this is not a hackathon story. It is a governance and benchmarking story dressed as a competition—and it lands precisely where the AI-RAN conversation has been weakest. The industry has spent two years proving that AI can improve a RAN. It has spent almost no time proving that an AI agent can be trusted to operate one. The GSMA challenge is an attempt to build the missing measurement layer before autonomous RAN claims harden into procurement assumptions.

The announcement, stripped of press-release varnish

The facts are straightforward. The GSMA’s Open Telco AI initiative has launched two competitions aimed at operators, AI researchers, developers and startups. TelcoRetrieve targets the accuracy problem: AI systems answering technical questions grounded in telecoms standards. The AI Agents for RAN Challenge targets the reasoning problem: agents that investigate network conditions, test possible interventions, and account for how changes ripple across the network.

The task list is unusually concrete for this kind of initiative: fault diagnosis, outage impact analysis, coverage and radio configuration, site placement and energy management. Participants work through Giulia, the open-source radio network simulator from UPV. The stated aim is to advance the AI capabilities needed for more autonomous radio network operations—improving service reliability, reducing operating costs, improving customer experience.

Judging will evaluate accuracy, efficiency, reasoning capability and security considerations. GSMA, ITU and IEEE will host a webinar on 18 November to introduce the challenge, explain how participants use Giulia to investigate scenarios and test interventions, and outline evaluation. The prize pool is €30,000.

That is the whole announcement. The significance is in what it implies.

Why this is the right problem, at the right moment

Walk through any operator’s current AI-RAN narrative and you will find the same gap. Vendors can now point to live demonstrations, field trials and commercial platform launches. What they cannot yet point to is a shared, reproducible way to answer a simple question: when an AI agent changes a radio parameter at 3 a.m., how do we know it reasoned correctly rather than got lucky?

That question is becoming commercially urgent. Operators are setting autonomy targets—AT&T, for instance, has publicly set an internal goal of reaching Level 4 network autonomy by 2030, and is moving its RAN, core, IP and optical layers toward open, disaggregated, programmable architectures. Verizon has talked about Level 4 cognitive automation as it deploys rApps on the Ericsson Intelligent Controller. Rakuten Symphony has claimed TM Forum Level 4 autonomy in AI-native Open RAN energy optimization. Level 4, in the TM Forum framing, means the network can operate and resolve problems itself within defined boundaries. The entire claim rests on the quality of the agent’s reasoning. Yet there is no industry-standard way to score that reasoning.

The GSMA challenge is, in effect, an attempt to build the exam. By anchoring tasks in a simulator rather than a live network, it creates a controlled environment where interventions can be tested, compared and scored—without risking a commercial service. That is the same logic that made open-source reference stacks and interoperability plugfests work for Open RAN. You cannot compare implementations if everyone is measuring against a different network.

Giulia is the quiet star of the announcement

The most under-discussed element here is the choice of simulator. Giulia, developed at UPV, is open source. That matters for three reasons.

First, reproducibility. An agent that scores well in a closed vendor simulator proves little. An agent that scores well in a public simulator can be re-run by a skeptical third party. If the industry wants AI agents in the RAN control loop, reproducibility is not a nice-to-have—it is the precondition for regulatory and internal-risk sign-off.

Second, accessibility. A €30,000 prize pool is not going to attract the largest vendors’ best teams. It is designed to attract exactly the population the industry has struggled to reach: AI researchers, startups and independent developers who do not have access to a Tier 1 operator’s test network. Open simulation lowers the barrier to entry in the same way open RAN interfaces lowered the barrier for radio vendors.

Third, and most strategically, it shifts the locus of competition. If the industry converges on shared simulation environments for evaluating agentic RAN behavior, the value migrates toward the quality of the agent and away from the exclusivity of the testbed. That is a structurally pro-Open RAN outcome, even though the challenge itself is not branded as an Open RAN initiative.

The four judging criteria are a roadmap of unsolved problems

Read the judging criteria as a diagnostic rather than a rubric. Each one maps to a known weakness in agentic telecom AI.

Accuracy is the TelcoRetrieve problem spilling over. Language models are fluent but not reliably grounded in standards documents. An agent that confidently misstates a 3GPP parameter is worse than no agent at all. Grounding answers in telecoms standards is the unglamorous prerequisite for everything else.

Efficiency reflects the operational reality that agents must run continuously, not once per demo. Inference cost at network scale is a real line item. An agent that solves a fault diagnosis beautifully but requires disproportionate compute per cell is not deployable.

Reasoning capability is the explicit target: does the agent investigate, form hypotheses, test interventions and account for network-wide effects? This is the difference between an automation script and an agent. It is also the hardest thing to evaluate, which is why a simulator-based task set is the right instrument.

Security considerations are the criterion most likely to be underweighted by entrants and most likely to matter to operators. An agent with tool access to radio configuration is an attack surface. The industry has already been reminded—through widely discussed agent breach incidents elsewhere in the AI ecosystem—that autonomous systems with tool permissions demand threat modeling from day one, not after deployment. Including security in the judging criteria signals that the organizers understand the stakes.

How this fits the broader AI-RAN buildout

The timing is not accidental. The last several months have produced a steady drumbeat of AI-RAN commercialization signals: T-Mobile’s claimed first AI-RAN call on a commercial network, Deutsche Telekom, BT Group and Elisa initiatives around Nokia’s CUDA-accelerated anyRAN software and NVIDIA’s Aerial RAN Computer, and NVIDIA’s count of nearly 20 operators with public AI-RAN collaborations. On the vendor side, Ericsson has been rolling out AI-native RAN automation through its Intelligent Controller and rApps with operators including Verizon and Batelco; ZTE has deployed a pre-commercial AIR RAN solution with Ethio Telecom, reporting an 11% increase in 4G capacity and a 40% gain in peak-hour user throughput.

All of that is real progress. But note what it measures: capacity, throughput, efficiency. These are outcomes of AI applied to RAN optimization—the “AI for RAN” half of the story. The harder half—agents that reason about the network and act on it—is where the evidence base is thinnest. The GSMA challenge is aimed squarely at that gap.

There is also a standards-politics dimension worth flagging. The lineup—GSMA, ETSI, IEEE, ITU, with Huawei as a collaborator—spans the institutional map of telecom. Getting these bodies to agree on a shared evaluation framework for agentic RAN behavior is a slow, political process. A competition is a pragmatic way to start building the evidence base before formal standardization. If the challenge produces a credible task set and scoring methodology, it becomes a de facto input to future standards work. That is how informal benchmarks become formal requirements.

The skeptical read

Three caveats deserve airtime.

First, simulation is not deployment. Giulia-based tasks cover fault diagnosis, outage analysis, coverage and radio configuration, site placement and energy management—a genuinely useful scope—but a simulator cannot fully capture the messiness of a live multi-vendor network with imperfect telemetry. A high score is necessary, not sufficient.

Second, a €30,000 prize pool is small. It will not move a large vendor’s roadmap. Its value is in seeding a community and, more importantly, in establishing a public benchmark that others can adopt and extend. Judge it as infrastructure, not as an incentive.

Third, the challenge does not by itself resolve the governance questions operators face. Who authorizes an agent to act? What are the guardrails? How is an agent’s action audited after the fact? Those are operator and standards-body problems. But a shared evaluation methodology is a prerequisite for answering them credibly—you cannot set guardrails around a capability you cannot measure.

What operators should do with this

For operators, the practical implication is to treat the challenge as a low-cost way to pressure-test internal assumptions. If your AI-native RAN roadmap assumes agentic automation by a certain date, you should be able to articulate how you will evaluate agent reasoning quality. The GSMA task set—fault diagnosis, outage impact, coverage and radio configuration, site placement, energy management—is a reasonable starting taxonomy for an internal evaluation program, whether or not you enter the competition.

For vendors, the message is subtler. A public, simulator-based benchmark for agentic RAN behavior is a double-edged sword: it creates a stage for differentiation, but it also creates a scoreboard that customers will use. Vendors that engage early can shape the methodology. Vendors that ignore it risk being measured by someone else’s rubric.

For the AI research community, this is one of the more accessible on-ramps into telecom yet offered. The combination of an open simulator, a concrete task list and a modest but real prize is designed to convert general-purpose agent research into telecom-specific capability. The 18 November webinar is the entry point.

The takeaway

The AI-RAN conversation has matured from “can AI improve the RAN?” to “can we trust AI to run the RAN?” The GSMA, ETSI, IEEE and ITU challenge is the first serious attempt to build the measurement layer that question requires. It is small, it is simulated, and it is imperfect. But it addresses the single biggest missing piece in the autonomous network story: a shared, reproducible way to know whether an agent’s reasoning is sound.

For Open RAN operators, the strategic read is encouraging. An open simulator, a public task set and multi-SDO backing push the industry toward verifiable, portable intelligence—the same principles that made disaggregated RAN worth pursuing in the first place. If the challenge succeeds, the next phase of AI-RAN competition will be about who builds the best reasoning, not who owns the testbed. That is a fight worth having.

Sources

Related Posts

Vodafone's Málaga Lab and the Silicon Fork Behind AI-RAN

Vodafone opens a multi-vendor AI-RAN hub in Málaga while VodafoneThree trials Nokia and NVIDIA — and Orange weighs GPU versus ASIC.

VodafoneThree's AI-RAN Lab Goes Live: Inside the Málaga Hub Where AI-for-RAN, AI-and-RAN, and AI-on-RAN Get Measured

VodafoneThree opens a Málaga AI-RAN hub with Nokia and NVIDIA, testing AI-for-RAN, AI-and-RAN and AI-on-RAN on shared compute ahead of 2027 field trials.