Daily Dose of 6G

3GPP has written down what an AI agent is

Standards bodies do not define things casually. TR 22.870 put a formal definition of “AI Agent” into a requirements document — and that definition is now propagating into SA2 architecture, CT3 protocol work, and a normative Stage 1 specification that freezes in March 2027.

Transcribed from 3GPP’s published Release 21 timeline (© 3GPP 2026). The highlighted row is where TS 22.270 — the normative home of these requirements — freezes.

The definition itself

3GPP’s own write-up of the 6G architecture work quotes SA1’s definition directly. An AI Agent is:

Read it as a specification writer would. Every clause is load-bearing. On behalf of another entity means the network’s counterparty may not be a human subscriber. Acquiring contextual information implies exposure interfaces. In collaboration with other AI Agents implies an agent-to-agent protocol, addressing and identity. None of that is decided by SA1 — but all of it is now owed by somebody, because the requirements document says the system shall support it.

Who wrote it, and where

TR 22.870, “Study on 6G Use Cases and Service Requirements”, ran as work item FS_6G_REQ under WID SP-241391 from 4 September 2024 to 3 March 2026. The portal names Xiaonan Shi (China Mobile) as rapporteur; 3GPP’s SA1 write-up records Jean Trakinat (T-Mobile USA) as co-rapporteur alongside — a deliberate pairing of a Chinese operator and a US operator on the release’s founding document.

The study item description was agreed at SA1#107 in Maastricht, August 2024, with the support of more than 90 companies — the most under-quoted statistic in Release 20, and not how every generation has started. SA1 met at #113 in Goa, 9–13 February 2026 weeks before the study closed, and the TR was approved at TSG#111 in Fukuoka, 9–13 March 2026.

The pen changes hands

Here is the detail worth noticing. The study is finished; the normative work is not the same document, not the same number, and not the same rapporteur.

NTT DOCOMO now holds the pen on normative Stage 1 — on top of RAN1 physical layer, SA2 system architecture and the parent FS_6G_Radio work item. That is an unusually concentrated set of seats for one operator, and it is worth watching whose requirements survive the study-to-spec transition.

The obligation propagates

SA1 requirements are cheap to write and expensive to honour. Three groups are currently paying:

• SA2 — TR 23.801-01, rapporteur Malla Reddy Sama (NTT DOCOMO). Eight work tasks, four of them under “beyond connectivity”: AI, ISAC, data frameworks, and computing service support coordinated across UE, core network and application server.
• CT3 — TR 29.832, rapporteur Apostolos Papageorgiou (Nokia Germany). Studying agent identity management, authentication and authorisation, dynamic networking and capability discovery, intent processing, task collaboration and multi-modal data transmission. Candidate protocols under analysis include Agent2Agent (A2A), Agent Network Protocol (ANP), AGNTCY and Model Context Protocol (MCP).
• SA3 — TR 33.801-01, rapporteur Todor Gamishev (Orange), which inherits the problem of authenticating an entity that is not a person and not a SIM.

The CT3 list is the concrete one. 3GPP is not inventing an agent protocol from nothing; it is evaluating four externally-developed ones and deciding whether to adopt, extend or replace. That is a different posture from how 3GPP has historically handled protocol design, and it is happening because the agent ecosystem moved faster than the release cycle.

Why this is the load-bearing edition

The radio studies decide how fast 6G goes; the agent thread decides what the network is for. A system treating software agents as first-class subscribers needs identity, authorisation, addressing, charging and lawful interception for entities that never held a phone — which is why TR 32.801-02 on charging and TR 33.801-06 on lawful interception are live studies rather than afterthoughts.

The caveat: none of this is settled. TS 22.270 is at v0.2.0 with nine months to Stage-1 freeze, and a requirement in a v0.2.0 draft is a proposal with good seats, not a commitment.