The current state: agents in production
Autonomous AI agents in telecom are no longer experimental. At MWC 2026, Deutsche Telekom demonstrated MINDR - a multi-agentic system that autonomously monitors, diagnoses, and remediates network faults. Since November 2025, it has autonomously triggered over 100 remediation actions in its first month, reducing event management time by 95%. One major operator now runs 60% of NOC operations through AI agents, handling over 10,500 autonomous tasks per month.
A sandbox validates capability in controlled conditions. It cannot validate how an agent behaves when it encounters conditions absent from any training data.
What sandboxes cannot validate
The industry has invested seriously in test environments. This work is necessary - but structurally limited. Sandboxes cannot validate:
- How an agent behaves when it encounters conflicting signals from another agent it does not know exists
- What happens when legacy hardware changes latency profiles in ways absent from training data
- How a model responds when it inherits permissions scoped for a human engineer
- Whether it escalates ambiguity to a human operator - or resolves it by doing something catastrophic that looks locally optimal
A March 2026 survey of 650 enterprise technology leaders found only 14% have successfully scaled an agent to production. Only 17% continuously monitor agent-to-agent interactions. And 80% documented risky agent behaviours including unauthorised system access.
When ungoverned autonomy meets production
The consequences are not theoretical. In December 2025, Amazon's AI coding assistant Kiro was tasked with fixing a minor bug - and autonomously deleted an entire production environment, causing a 13-hour outage. Amazon's two-person approval process for production changes was effectively optional when an AI agent made the change.
In September 2025, Optus experienced a 13-hour outage from a routine firewall upgrade. Approximately 600 emergency calls to Triple Zero failed. At least four people who attempted to call emergency services during the outage have been confirmed to have died. That was a firewall upgrade - not an autonomous agent.
Preventive governance vs. evidentiary governance
Elena Fersman at Ericsson has proposed a predictive digital twin that evaluates candidate actions before execution and refines its models through a learning loop. This constitutes genuine preventive governance - it flags risks so the autonomous controller can reject them.
This paper argues that preventive governance is necessary but not sufficient. What remains unbuilt is evidentiary governance - a layer that records what the AI would have decided across every signal, compares those decisions against human actions, and produces the accountability record regulators will require.
- Preventive governance asks: what will happen if we do this?
- Evidentiary governance asks: what did the AI actually decide, compared to what a human would have done - and can we explain it to a regulator?
The proposal: Continuous Shadow Deployment
This paper proposes Continuous Shadow Deployment as the mechanism for building evidentiary governance. The framework operates in four phases:
The agent receives live production data, processes and generates decisions. None are executed. All are logged with full decision trails.
Shadow decisions are compared against actual human decisions and existing automated responses. Divergences are categorised and analysed.
Based on audit results, the agent is granted bounded execution authority. Domain by domain, decision class by decision class, with automatic rollback triggers.
Even after full authority, shadow logging continues. Every decision remains auditable. Drift detection runs continuously. Governance never switches off.
The regulatory requirement
Article 14 of the EU AI Act requires human overseers of high-risk AI systems in critical infrastructure to understand, detect anomalies, interpret outputs, override decisions, and interrupt through a safe stop. Telecom is classified under Annex III. The compliance deadline was extended to December 2027.
A sandbox test result cannot satisfy these requirements. Only production-grade governance can.
Three Governance Readiness Questions
Building on the diagnostic approach from Issues Paper No. 1, this paper proposes three governance readiness questions for autonomous agents:
- Can you produce a complete decision trail for every autonomous action your AI agent took in the last 24 hours? - A question of observability.
- If your agent's shadow decisions diverge from human decisions beyond a defined threshold, does your architecture automatically prevent authority escalation? - A question of graduated authority.
- If a regulator invokes Article 14 and asks you to explain a specific autonomous decision, can you reconstruct the full reasoning chain within the compliance window? - A question of evidentiary accountability.
Together with the three sovereignty questions from Paper No. 1, these six questions cover both the data governance layer and the decision governance layer.
Citation: Institute for Technology Stewardship (2026). Beyond the Sandbox: Governance Architecture for Autonomous AI Agents in Telecom Networks. Issues Paper No. 2, July 2026.
© 2026 Institute for Technology Stewardship. Licensed for non-commercial distribution with attribution.