Security is no longer reactive. It has to predict, prove and act

Why autonomous networks need a semantic agent protocol

Autonomous network deployments are increasingly constrained by a trust and interoperability gap. The models exist, and the agents exist, but many deployments still stall at Level 3 because agents from different vendors cannot reliably coordinate structured work, share operational meaning, or follow an operator’s own processes in production.

The Catalyst project, Agent Fabric: A2A-T Runtime Phase III, addresses that missing layer. The project showcased at DTW Ignite in June 2026 and was recognized as Outstanding Catalyst for AI & Automation, awarded to the project that demonstrates the most advanced use of AI & Data, GenAI, Agentic frameworks or Automation to transform telecom operations, business models, or customer engagement

The commercial risk is clear: without a standard way to coordinate, govern and audit autonomous decisions, agents that work in lab environments cannot be trusted in real network operations. There is no consistent evidence trail that survives a handoff, and no common way to prove that a decision was made correctly.

For operators pursuing higher levels of autonomy, this blocks the move from isolated automation to trusted, multi-vendor orchestration across domains.

A2A-T runtime for governed multi-vendor agent collaboration

Agent Fabric delivers production infrastructure built on A2A-T, a telecom-grade extension of the Linux Foundation A2A protocol. A2A-T gives agents a common language for structured tasks, real-time events and capability negotiation, so they can collaborate across vendors and domains in a deterministic way. Linux Foundation A2A lets agents communicate. A2A-T adds the telecom meaning they communicate about, so a task issued by one vendor agent is interpreted the same way by the next. It is an extension, not a fork. It uses the Linux Foundation A2A extension mechanism as intended, leaves the core protocol untouched and stays backward compatible. Three extensions carry the telecom semantics: Task-T structures intent as governed, standards-anchored work, Event-T handles proactive cross-domain event subscriptions, and Negotiation-T resolves target and feasibility before execution. Grounding agent exchanges in this structure improve accuracy, lowers token cost, and reduce exposure to prompt injection.

The solution goes beyond the wire protocol. It binds agent activity to TM Forum HVS tasks, operator processes and real operational scenarios, enabling agents to understand telecom context rather than simply exchange messages. Semantic discovery allows any agent to find the right partner agent based on capabilities and meaning, not just preconfigured integrations. Trusted Gateway component ensures that only allowed communication activities among agents will be executed by providing a protection mechanism from free-text commands or data injection attempts.

Trust in production it about “evidence-based confidence.” Operators govern the system at design time. Approved processes are registered as executable playbooks, and agents execute those playbooks rather than inventing behavior. This is human-on-the-loop governance, not human-in-the-loop approval of every action. Every cross-vendor exchange is then recorded in an immutable audit log, attributed to the agent that acted and replayable after the fact, so an operator can show exactly how an outcome was reached. The agents are responsible for execution. The operator remains accountable for the result. That is the “protocol to proof” arc the Catalyst is built to close: from a structured protocol, through a governed runtime, to proof an operator and a regulator can stand behind. This is how trusted is earned – slowly and pa

The team both draws on and advances TM Forum assets. It is driving the standards of work for the A2A Protocol Telecom Extension (aka A2A-T) through IG1453 and IG1453A. It has also developed a fabric Knowledge Graph and a SID ontology that make TM Forum knowledge machine-consumable for agents, work that is helping shape TM Forum's wider ontology direction. Its biggest contribution is working code. Through the Linux Foundation OpenAN Project, the team has open-sourced running components of the fabric, including the Agent Registry, the Orchestration Centre and the A2A-T protocol itself, so operators and vendors can take a working implementation rather than a specification alone

From faster fault resolution to trusted Level 4 autonomy

The Catalyst demonstrated one fabric across three high-value scenarios: cross-domain service fault management, enterprise complaint handling, and enterprise service fulfilment; each executed end to end without human handoff. The potential impact is significant. The project team reports that autonomous multi-vendor agent chains could reduce fault resolution time by up to 80% by eliminating human handoffs, while cutting NOC effort on cross-domain diagnosis by up to 90%.

In enterprise service fulfilment, the same approach could compress delivery cycles from weeks to hours. For SLA assurance, the project points to zero-touch resolution in which a breach can be detected, the root cause identified, the customer credit issued, and the customer notified through a governed autonomous workflow.

By making A2A Telecom Extension an open standard (IG1453) and open-sourcing the actual protocol extension code and the runtime under OpenAN, the Catalyst creates a shared foundation that any operator or vendor can build on, reducing lock-in and helping the industry move toward trusted, federated autonomous networks.

Pedro Garcia Parra, Director of Autonomous Networks at Telefónica, said: “At Telefónica, the unsolved challenge in Autonomous Networks is not building individual agents — it is keeping them working together safely at scale. By extending A2A with semantics and a governance fabric to enable trust, this Catalyst makes production-level L4 autonomy achievable today.”