Telecom networks are already highly automated. Configuration templates deploy infrastructure, assurance systems raise alarms, policies trigger closed loops, and AI/ML models optimize individual functions. Yet automation is not autonomy.
Most operational systems remain collections of bounded automations. Humans still formulate the workflow, select the algorithm, reconcile competing business objectives, approve risky changes, and redesign the system when conditions move beyond its original assumptions. Adding more scripts or a larger AI model does not close this gap.
The real challenge is to govern a changing network as a coherent system: understand what stakeholders want, determine what is feasible, choose or generate the right method, verify its consequences, act through the correct control horizon, observe the result, and revise the plan.
BubbleRAN definition: An autonomous network is a self-governing communications system that continuously translates stakeholder intent and live network conditions into safe, verifiable action - reconfiguring its services, APIs, control logic, and resources; detecting and resolving faults; and improving its capabilities across time and space without routine human intervention.
Autonomy is therefore not a single AI model, dashboard, agent, or automated script. It is a system property. It emerges when sensing, reasoning, optimization, validation, actuation, governance, and learning operate as one continuous loop.
Automation is not yet autonomy
Industry frameworks provide essential foundations. TM Forum measures autonomy by operational scenario and maturity level. ETSI ZSM advances end-to-end zero-touch management. O-RAN creates open control surfaces through the SMO, Non-RT RIC, Near-RT RIC, rApps, xApps, and emerging dApps.
These frameworks explain how networks progress from manual operations toward increasingly closed-loop behavior. BubbleRAN extends this progression with a system-level question:
Can the network turn changing intent into continuously verified outcomes across domains and timescales, while preserving operator authority, policy compliance, sustainability and the ability to recover?
A practical definition
An autonomous network turns intent into continuously verified outcomes, across domains and timescales, without routine human intervention.
Four words in this definition matter:
- Intent expresses the desired outcome and its constraints, rather than a specifc sequence of commands.
- Continuously means the network adapts as topology, traffic, radio conditions, faults, energy budgets, and stakeholder priorities change.
- Verified means candidate actions are checked against policies, feasibility constraints, a network digital twin, runtime monitors, and rollback conditions.
- Outcomes shift the unit of autonomy from isolated tasks to service and business objectives such as coverage, capacity, latency, resilience, efficiency, trust, and user experience.
The Figure shows how BubbleRAN realizes the autonomous network control loop. Stakeholder intent is interpreted by cooperating agents and optimizer-grounded AI pipelines. Candidate actions are verified through a Network Digital Twin before being deployed through open SMO/RIC control loops. Live telemetry closes the learning cycle, while governance surrounds every stage. Self-Synthesizing Networks represent a future research direction in which networks compose and evolve their own capabilities.
Connectivity on an offshore platform hundreds of kilometers from shore can’t rely on constant manual intervention, and every optimization affects safety, operations, and energy efficiency. This is where autonomous networks become more than an efficiency upgrade, they become an operational necessity.
How BubbleRAN extends the current industry model
| Lens | Industry emphasis today | BubbleRAN extension |
|---|---|---|
| Autonomy target | Zero-touch, self-managing operations assessed by scenario and maturity level | A system-level loop that can also adapt its control logic, APIs, and capabilities |
| Intent | Declarative goals translated into policies and workflows | A negotiable, living contract among stakeholders, constrained by feasibility, policy, and trust |
| AI role | Analytics, prediction, policy, and function-specific optimization | Multi-agent reasoning coupled symbiotically with precise optimizers, tools, and managed AI pipelines |
| Safety | Policies, assurance, governance, and staged closed loops | Pre-validation in a live network digital twin, bounded execution, runtime monitoring, auditability, and rollback |
| Control | Domain controllers and standardized loops at defined timescales | Agents coordinate rApps and xApps - and, as the architecture evolves, dApps - across control horizons |
| Evolution | Automation improves as engineers add models, rules, and workflows | A path toward networks that can synthesize, validate, deploy, and retire capabilities as conditions change |
The autonomy stack: one loop, several cooperating systems
BubbleRAN’s roadmap is concrete because each technology maps to a necessary part of the operational loop.
1. Sense: establish a trustworthy operational data plane
Autonomy starts with a consistent and traceable view of the network. Opti-Sphere connects to existing EMS, NMS, SMO, RIC, RAN, Core, and infrastructure environments. It normalizes KPIs, logs, alarms, configuration, topology, infrastructure usage, and - where available - energy signals into operational context.
This layer is not merely a data collector. It must preserve identity, time, lineage, scope, and quality. Without those properties, even a sophisticated agent is reasoning over noise, stale context, or incompatible measurements.
2. Understand intent: create a human interface to outcomes
Operators and service owners should be able to express objectives in their own language: protect an SLA, improve uplink reliability, reduce energy use, prioritize an emergency service, or rebalance resources for a temporary event.
AI capabilities of Opti-Sphere demonstrates how natural-language intent can be routed through specialized agents for observability, planning, configuration, deployment, and control. The interface is not cosmetic. It is where objectives, constraints, approvals, explanations, accountability, and operator authority meet.
3. Reason together: use multi-agent systems, not a single omnipotent agent
No single agent should own every decision. Specialized agents can plan, monitor, deploy, validate, retrieve knowledge, manage a twin, inspect APIs, and coordinate control actions.
MX-AI implements a cooperating agent graph in the SMO layer and connects it to a live Open RAN through R1 and E2. This division of responsibility makes complex behavior more composable, inspectable, and governable. It also makes it possible to replace or improve one capability without rebuilding the entire system.
4. Optimize precisely: combine agents with optimizers and managed AI pipelines
Language models are valuable for interpreting goals, selecting tools, explaining choices, and adapting workflows. They are not, by themselves, reliable numerical optimizers.
BubbleRAN’s Symbiotic Agents pair agentic reasoning with mathematical optimization and control. An agent can generate, choose, configure, and tune an optimizer; the optimizer supplies feasibility, bounded uncertainty, and numerical precision; and the agent supervises the evolving task.
In published experiments, this design reduced decision errors fivefold compared with standalone LLM agents. It also demonstrated flexible SLA and resource negotiation with approximately 44% lower RAN over-utilization.
The same principle applies to AI/ML pipelines. Data selection, feature preparation, model or algorithm choice, training, validation, deployment, monitoring, drift detection, and replacement must become governed lifecycle actions - not isolated data-science projects.
5. Imagine safely: verify consequential actions in the Network Digital Twin
The Digital Network Digital twin included in Opti-Sphere is the proving ground between intention and impact. It can mirror a production network at the scope of a site, slice, region, or larger system, with emulated UEs in the loop, synchronized or frozen state, parallel replicas, and open integration with the SMO and RIC.
Candidate actions can be tested against realistic alternatives before touching production. For autonomous operation, the twin becomes an always-available validation service:
- What is likely to happen if configuration changes?
- Which policy option best satisfies the intent?
- What could fail based on frequent root causes?
- How confident is the system to guarantee the SLA?
- Which guardrails and rollback thresholds should be armed?
A network cannot responsibly move toward zero routine intervention unless it can rehearse consequential changes and compare outcomes before acting.
6. Act at the right timescale through open control loops
Reasoning has value only if the system can act. RIC-Sphere provides Non-RT and Near-RT control, while SMO-Sphere provides intent-driven lifecycle automation from Day 0 to Day 2+.
Agents can deploy and coordinate:
- rApps for longer-horizon policy, assurance, data, and AI/ML orchestration,
- xApps for faster RAN monitoring and control, and
- dApps for control close to the CU/DU.
Open interfaces - including R1, A1, E2, and operational APIs - turn autonomy into an interoperable control system rather than a proprietary AI overlay.
7. Govern execution: progress through a ladder of trust
Operators do not need to jump from dashboards to unbounded autonomy. Opti-Sphere supports a practical progression:
- Recommend-only: rank actions and explain the evidence; the system does not execute.
- Human-approve: produce an approval-ready plan with expected impact, constraints, and an audit trail.
- Bounded closed loop: execute autonomously inside explicit policies, monitors, staged rollout, and rollback rules.
Trust is earned by making decisions observable, validating them in the twin, automating a well-understood envelope, and expanding that envelope as evidence accumulates.
A 2026 deployment with Telenor on a Color Line vessel illustrates this architecture in a difficult radio environment. Opti-Sphere combines AI agents and algorithms, continuously optimizes uplink power under operator-defined constraints, and pre-validates candidate configuration changes through a real-time digital twin.
8. Reconcile interests: autonomy must negotiate what “optimal” means
A network rarely optimizes one objective for one owner. Operators, tenants, vertical applications, infrastructure providers, regulators, and users can have conflicting goals.
AGORAN reframes resource control as a governed marketplace. Stakeholder negotiation agents and a mediator generate feasible, Pareto-oriented offers. Dedicated governance branches handle compliance, situational awareness, trust, and arbitration. The resulting consensus intent is deployed through Open and AI-RAN controllers.
In a private 5G testbed, the published evaluation reported:
- 37% higher eMBB throughput,
- 73% lower URLLC latency, and
- 8.3% lower PRB usage than a static baseline.
The broader point is architectural: an autonomous network must not merely optimize. It must continuously reconcile what different stakeholders mean by an acceptable and trustworthy outcome.
Beyond self-optimization: Self-Synthesizing Networks
Today’s automation executes capabilities that engineers have already designed. Even advanced closed loops usually select among known policies, models, applications, or parameter ranges.
The next research step is the Self-Synthesizing Network (SSN): a network that can compose, generate, validate, deploy, evolve, and retire capabilities as conditions change across time and space.
In an SSN, the question is no longer only, “Which configuration should I apply?” It becomes:
- Which capability is missing for this new situation?
- Can existing RAN functions, agents, data pipelines, models, optimizers, rApps, or xApps be composed to provide it?
- Must a new control workflow, API adapter, model, or application be generated?
- Can the synthesized capability be verified in the NDT, constrained by policy, deployed progressively, observed, and rolled back?
- Should it be localized to one site or slice, propagated across a region, or retired when the context disappears?
SSN is a forward-looking research direction, not a claim of fully autonomous commercial operation today. It sharpens the destination: the network should eventually evolve not only its state, but also the machinery through which it senses, reasons, and acts.
A concrete road to autonomy
| Stage | Operational milestone | BubbleRAN building blocks |
|---|---|---|
| 1 - Observe | Unify data, topology, configuration, and service context; establish lineage, baselines, and explainable insights | Opti-Sphere L1/L2; SMO/RIC observability |
| 2 - Recommend | Turn intent into ranked, feasible actions with confidence and evidence | Opti-Sphere; MX-AI; optimization primitives |
| 3 - Verify | Test candidate changes and failure modes in synchronized, scoped digital twins | MX-DT; Digital Twin Agent |
| 4 - Execute safely | Deploy rApps/xApps and apply staged changes inside explicit guardrails, monitors, and rollback rules | SMO-Sphere; RIC-Sphere; Opti-Sphere bounded loops |
| 5 - Coordinate | Use specialized and symbiotic agents to select tools and algorithms across domains and timescales | MX-AI; Symbiotic Agents; managed AI pipelines |
| 6 - Negotiate | Reconcile multiple stakeholder intents, policies, trust, and resource constraints | AGORAN |
| 7 - Synthesize | Generate and evolve capabilities, validate them, and place them where and when they are needed | SSN research direction |
What “no human intervention” should mean
The strongest definition of autonomy remains a network that needs no routine human intervention. But no human intervention must not mean no human authority.
Humans define governance, access rights, risk appetite, legal obligations, safety policies, and the boundaries within which autonomy operates. They must be able to inspect, override, and change those boundaries.
Inside that governed envelope, the network should carry the operational burden itself: observe, diagnose, negotiate, plan, validate, execute, learn, and recover. The goal is not to remove people from the network. It is to move them from repetitive control of individual elements to deliberate governance of outcomes.
The road is already under construction
BubbleRAN’s approach connects research ambition to deployable systems:
- MX-PDK and the Sphere platforms provide the cloud-native, O-RAN-compliant substrate.
- SMO-Sphere and RIC-Sphere expose lifecycle and control loops.
- MX-AI supplies agentic observability, reasoning, and intent-driven control.
- MX-DT provides a safe, high-fidelity proving ground.
- Opti-Sphere turns operational data into governed recommendations and bounded automation.
- Symbiotic Agents ground reasoning in optimization.
- AGORAN introduces stakeholder negotiation and trust.
- SSN extends the vision from self-optimizing networks toward networks that synthesize the capabilities they need.
Autonomy is reached when the network can change itself without losing control of why it changed, who it serves, what it risks, and how it can recover.
Explore BubbleRAN’s autonomous networking building blocks and request a demonstration
Further reading
- BubbleRAN Opti-Sphere
- BubbleRAN MX-PDK Campus Network
- Agentic AI for Open RAN: Design, Integration, and Field Validation in a Multi-Agent Framework
- Symbiotic Agents: A Novel Paradigm for Trustworthy AGI-driven Networks
- AGORAN: An Agentic Open Marketplace for 6G RAN Automation
- Agentic AI Optimizing Private 5G Performance Live at Sea
- TM Forum Autonomous Network Mission
- ETSI Zero-touch network and Service Management
- O-RAN Digital Twin RAN: Key Enablers