Networks for the AI supercycle need autonomy.
This shift has a direct implication for operation: networks can no longer rely on static planning cycles and must adapt continuously, detecting, deciding, and executing adjustments through closed loops without human approval. This is unavoidable because changes occur faster than human operational cycles, the complexity of modern networks makes manual intervention impractical, and the effects spread across multiple domains, so isolated interventions can amplify instability rather than contain it.
Although network autonomy has been discussed for years, a clear framework for coordinating end-to-end decisions and governing them in a verifiable, controlled, and reversible way was still lacking. Without this foundation, network operation becomes more complex, as autonomy leads to more control loops, shorter decision times, and a greater impact of errors.
Coordination and Trust in Network Autonomy:
Deploying autonomy in separate domains is insufficient. If radio access, core, IP access, and transport apply independent AI loops to the same network states, systems can conflict, leading to oscillations or contradictory decisions. Therefore, autonomy cannot be isolated; it requires system-level coordination to arbitrate conflicts and ensure that intent is consistently translated across all domains. Without it, autonomy becomes fragmented and a structural risk.
Furthermore, trust in autonomy depends on transparency and governance. Operators must be able to see what changes were made and why, know which models were active, limit the frequency of changes, and safely revert actions. Congestion in the transport network can trigger adjustments between radio access and the core that do not converge, whereas coordinated autonomy ensures the network converges rather than oscillates.
The Glass Box Imperative
: To be deployed at operator scale, autonomy must be verifiable, controllable, and reversible. You must explain your actions, operate within explicit safety limits, and allow reversibility as standard practice. This is what we call a glass box.
Two realities reinforce this imperative:
1. Transition to agent-based automation: Across all industries, organizations are moving from AI-assisted automation to agents that execute multi-step workflows. In networking, agent-based automation is the natural model for continuous adaptation, but it requires a higher level of accountability.
2. The gravity well of intelligence: The primary source of technology dependency is shifting from hardware to the intelligence layer: models, policies, and context graphs. Monolithic, proprietary platforms can generate long-term dependency.
In networks with existing infrastructure, autonomy is not achieved through complete replacement. Multigenerational networks already combine domain controllers, policy systems, and operational processes from various vendors. Practical implementation is incremental: first, high-value closed loops in advisor mode; then limited actuation validating observability, provenance, and reversibility; and finally, expansion as trust is gained. This leads to a hybrid agent model: Local domain agents execute rapid decisions within strict boundaries. At the same time, an end-to-end governance layer coordinates objectives, resolves conflicts, and controls permissions and deployments, thus achieving unified logical coordination with distributed execution.
Architecture: Intent, Agents, Governance, and Operational Structure.
The architectural challenge lies in structuring intelligence so that autonomy extends from beginning to end without generating instability, opacity, or dependency. One way to frame this is as an operational structure that connects four elements: what the operator wants, what the system knows, what the system is allowed to do, and how it demonstrates that it acted correctly. These elements transform autonomy from a set of optimizers into a governable system. In this context, an agent is software that translates intent into action, discovering the context and invoking operational tools under permissions and security boundaries. Agents are classified as observers, advisors, actuators, coordinators, and lifecycle agents, and together they illustrate how intent, agents, governance, and execution are integrated into a system.
1. Intent: defines outcomes, latency, reliability, and energy trade-offs, and functions as a common language across domains.
2. Agent-based automation: translates intent into sequences of coordinated actions, with authorization and policy boundaries.
3. AI Governance: A distributed coordination plane that arbitrates conflicts and enforces global policy, preventing local logic from overriding the service's intent.
4. Bounded Execution: Ensures operational safety through explicit limits on scope, frequency, and range.
The end result is a unified operational structure where automation reasons about interrelated outcomes rather than local key performance indicators (KPIs).
Five properties that enable glass-box autonomy.
For automation to be reliable and governable, the system must meet five essential properties:
• Observability: Every action must be visible; it must be possible to track what changed and when.
• Provenance: The system must record which model or policy was responsible for each decision.
• Traceability: It must be documented what signals and context motivated the action.
• Safety controls: Clear boundaries must be established to prevent unintended effects, including the ability to revert changes.
• Auditability: Every intervention must be attributable and verifiable, allowing for review of decisions and compliance with regulatory requirements.
These properties are essential for trusting autonomy, analyzing past events, complying with regulations, and responding responsibly to incidents.
Lifecycle, Governance, and Openness of Autonomy:
The lifecycle of autonomy encompasses much more than code, including AI models, policies, agent flows, and data governance. Automation requires incorporating continuous integration (CI) for testing and validation before deployment, and continuous deployment (CD) to implement decisions and agents with impact tracking. Data traceability and access control are critical, as every decision depends on what information the models processed, and without this, transparency and regulatory compliance are not guaranteed.
The risk of technological dependence no longer lies in the hardware, but in the network's intelligence layer: models, context graphs, APIs, and agent flows. To avoid technological lock-in, open and governed interfaces are needed to enable interoperability, as well as ensuring the portability of automation, so that flows, models, or agents can move between platforms without being locked by specific vendors.
Autonomy is inevitable, instability is not.
AI is embedding intelligence into every layer of the network. The glass box is the limit that remains, the only way to scale autonomy without sacrificing the determinism and trust necessary for essential infrastructure. The only real risk is moving too slowly while autonomy develops faster than our capacity to manage it.
Authors: Oğuz Sunay, CTO Fellow of AI, and Pallavi Mahajan, Director of Technology and AI at Nokia
