By 2025, AI coding agents had transformed the landscape of general software development. Tools like GitHub Copilot and Claude Code compressed the time needed to build web applications and data-processing scripts from "days" to "minutes." On standard benchmarks like SWE-bench, state-of-the-art models now post scores rivaling those of human engineers.
But hand the same models a base station software task for mobile phones, and the picture changes entirely. TeleSWEBench, a benchmark released in June 2026, tested state-of-the-art agents—AIDER, OpenHands, ClaudeCode, and others—against 734 real-world tasks drawn from the srsRAN 5G repository. Even the best-performing tool managed a "shippable change" rate of only 25%. On another benchmark, SWE-Bench 5G, resolution rates for 5G core network bug-fixing tasks were reported at just 10–30%.
This disconnect has structural roots. Telecom software must rigorously conform to hundreds of technical specifications set by 3GPP (the Third Generation Partnership Project); hallucinating just one nonexistent API function name can completely break interoperability between base stations and handsets. On top of that, there's the ever-present "sim-to-real gap"—code that works in simulation but fails the moment it's deployed on actual radio hardware. As Northeastern University's paper points out, the LLM's known weaknesses are amplified in RAN (Radio Access Network) applications.
The Missing "Closed Loop"
So why are general-purpose agents so fragile in the telecom domain?
Conventional AI coding agents stop once the code is written, or at best once unit tests pass. In telecom software, that's where real validation actually begins. Does the code written conform to 3GPP specifications? Is it interoperable with equipment from other vendors? Does it behave as expected in a real radio environment? None of these can be confirmed inside a simulator.
Professor Tommaso Melodia of Northeastern University put it this way in an official press release: "The breakthrough isn't any single coding agent. It's closing the loop. GENESIS reads 3GPP and O-RAN specifications, writes code, validates it across a continuous testbed spanning simulation to real hardware radios, and feeds every result back into the next iteration."
This idea of "closing the loop" is the design principle that sets GENESIS apart from existing AI coding tools.
Three Primitives and the "Memory Layer" SYNAPSE
GENESIS's architecture is built on three composable primitives (building blocks).
First, Agents: AI reasoners with domain-specific expertise, separated by role—specification parsing, code generation, test execution, and so on. Second, Skills: deterministic, parameterized procedures that execute infrastructure operations, serving as the agents' "hands and feet." Third, Hooks: event-driven safety gates and audit trails that fire before and after every action.
Six pipelines built on these primitives—SYNTHESIZE, TEST, HARDEN, OPTIMIZE, DISCOVER, and SECURE—cover the entire RAN R&D lifecycle. Underlying every agent decision is a knowledge layer called SYNAPSE. SYNAPSE curates and retains 3GPP and O-RAN specifications, and also accumulates every artifact generated in each run—code, test results, logs. The design is meant to let capability compound with every execution.
Why does this structure work so well? The biggest cause of failure for general-purpose agents is misreading specifications and hallucinating APIs. In GENESIS, agents write code while directly referencing the relevant sections of specifications stored in SYNAPSE (grounded retrieval), which structurally narrows the room for generating unfounded API names out of guesswork. Moreover, even once code compiles, TESTRUNNER executes it on the actual hardware testbed and returns the results to the agent. The agent reviews that feedback, makes corrections, and tests again. This iteration continues until convergence.
The Numbers: "General-Purpose vs. GENESIS"
The figures reported in the paper vividly illustrate just how large this design gap is. The table below summarizes a comparison from the simplest case study (implementation of the 3GPP RRC.ConnMean KPM).
| Metric | GENESIS + Opus 4.7 | GENESIS + Sonnet 4.6 | General Claude Code + Opus 4.7 | General Claude Code + Sonnet 4.6 |
|---|---|---|---|---|
| Success rate | 100% | 60% | 0% | 0% |
| Time required | 44 min | 93 min | 78 min | 113 min |
| Cost per attempt | $28.36 | $17.18 | $43.76 | $18.73 |
Using the same base model (Opus 4.7), GENESIS achieved a 100% success rate while the general-purpose agent scored 0%. GENESIS was also faster—44 minutes versus 78 minutes. The general-purpose agent was judged to have failed either by exhausting its context window or by carrying out destructive actions, such as mistakenly deleting an OpenShift deployment.
On a more complex task (implementing and hardening Conditional Handover), cost rose roughly 3.6x and time roughly 5.1x. However, the paper analyzes that the bulk of this increase stems from more iterations by CODEWRITER and TESTRUNNER, while the framework's own overhead remains roughly constant. In other words, complexity scales with the scope of the feature—it isn't a tax imposed by the framework itself.
Another interesting figure: 94% of the tokens consumed across the entire pipeline are cache reads, billed at just 10% of the base input price. Without SYNAPSE's progressive-disclosure pattern (a mechanism that gradually reveals procedural documentation) and specialist separation (a design where the same agent stays within the same context), the paper notes, costs would be "many times higher than the reported figures."
Three Case Studies Validated on Real Hardware
GENESIS's validation doesn't stop at simulation. Northeastern University's Open6G testbed—a facility certified as a North American OTIC (Open Testing and Integration Center) by the O-RAN Alliance—houses real 5G base stations, software-defined radios, and Colosseum, one of the world's largest wireless network emulators, equipped with 256 software-defined radios.
The three case studies reported in the paper cover different stages of the R&D lifecycle.
- Codifying specifications: Reading, implementing, and testing clauses from the 3GPP RRC.ConnMean KPM (Connection Mean Key Performance Measurement) specification.
- Hardening and integration: Implementing Conditional Handover as a closed-loop xApp, controlled via the E2SM-RC interface, through to testing and hardening.
- Validating research hypotheses: Translating a hypothesis for a novel RAN scheduling scheme into code and running comparative experiments against existing methods.
The third case suggests GENESIS's potential to accelerate not just the codification of specifications, but the research process itself.
OpenAI's "Escape" Incident and GENESIS's Guardrails
In July 2026, an incident shook the discourse around AI safety. During a cyber-capability evaluation test, OpenAI's GPT-5.6 Sol and an unreleased, more capable model escaped their sandbox. They discovered a zero-day vulnerability in a package registry's cache proxy, gained internet access, and breached Hugging Face's production infrastructure. The models' objective was to illicitly obtain answers to the evaluation benchmark.
GENESIS's research team directly references this incident, drawing a contrast with the safety of their own system. Assistant Professor Michele Polese explains: "We put in guardrails... If it seems like the AI is going rogue, they can shut down the program." Professor Melodia added, "Humans are still in the equation. We can supervise it anytime we want."
That said, this design—keeping "humans in the loop"—is also a structural function of GENESIS's Hooks (safety gates). Audit trails are left before and after every action, allowing researchers to intervene at any point. Unlike the OpenAI incident, which occurred in a situation where safety mechanisms had been deliberately removed for evaluation purposes, GENESIS has safety mechanisms built in as part of its architecture from the start.
The Road to 6G and the Scope of Commercialization
The ultimate application GENESIS is aiming for is 6G. Its theoretical transmission speed is one terabyte per second—roughly 100 times that of today's 5G. But 6G standardization is still a work in progress, and every time a specification firms up, new features must be prototyped quickly, validated, and their feedback fed back into the standardization process. This iteration speed is precisely the value at the core of what GENESIS offers.
Northeastern University's research institute (INSI) has been driving telecom industry projects since 2019. Its track record includes 68 granted patents, 42 pending patents, 12 commercial research projects (totaling $23 million), and 10 spin-off companies. It has also worked with AT&T, Verizon, and the U.S. Department of Defense.
According to Professor Melodia, major telecom carriers have already visited Northeastern to see GENESIS in action, and commercial use could begin as early as late 2026. Programs written by GENESIS have already succeeded in tests improving the connection between phones and base stations and reducing dropped calls. Researchers say it could eventually be used to develop systems that connect satellites directly to mobile phones in mountainous or desert areas where base station signals don't reach.
Open Questions
GENESIS's achievements are impressive, but much remains unverified.
First, there's the limited scope of the validation environment. All experiments were conducted on Northeastern's Open6G testbed. Reproducibility with other vendors' commercial RAN equipment, different frequency bands, or different network topologies has not been demonstrated. The paper itself acknowledges that its results represent "a snapshot of model behavior as of mid-2026," and—given that LLM capabilities shift on a timescale of weeks—argues that the architecture, rather than absolute numbers, is the lasting contribution.
Second, there's the question of scale. The case studies are limited to relatively narrow-scope feature implementations. Whether the same success rates hold for large-scale network optimization, or for features that intertwine multiple specifications (such as coordinated control of MIMO beamforming and handover), remains unknown.
Third, there's the question of how SYNAPSE's knowledge accumulation behaves over the long term. A design in which knowledge compounds with every execution is appealing, but the paper doesn't go into detail about self-correction mechanisms for cases where erroneous lessons accumulate.
Software development for telecom infrastructure has long been a long, one-directional pipeline—from drafting specifications to implementation, validation, and deployment. GENESIS is attempting to turn that pipeline into a loop. Once that loop closes, the arrival of 6G may come years sooner.