← Archive · 6G Tracker
 

Daily Dose of 6G

Edition 15 5 September 2026 Series: The architecture study

AI for the network, and the network for AI

The other half of TR 23.801-01 is four work tasks and five key issues. One of them has a live drafting folder holding twenty-eight files, filed the week SA2 met in Prague, with titles about intent, tool invocation and multi-agent communication. It is the first August activity this newsletter has been able to see.

Transcribed from 3GPP’s published Release 21 timeline (© 3GPP 2026). The highlight stays where Edition 14 put it: TR 23.801-01 is a Stage 2 document, and Rel-21 SA Stage 2 complete is the milestone everything below has to reach.

Four tasks, five key issues

Edition 14 took the Connectivity half of TR 23.801-01 — “Study on Architecture for 6G System; Stage 2”, specification rapporteur Stefan Rommer of Ericsson LM, work-item rapporteur Malla Reddy Sama of NTT DOCOMO. This is the other half: the four tasks Tao Sun, TSG SA Vice Chair, China Mobile, describes on 3gpp.org as moving “beyond connectivity through AI, ISAC, data frameworks and support of computing”. Between them they own five of the study’s twenty-four key issues.

TaskKey issueWhat it asks
WT#3_AIKI#18AI for 6G Architecture — how to “support and enable use of AI in 6G (e.g. AI agent, AI framework)”
WT#3_AIKI#196G Network for AI — agents on handsets discovering each other, inter-agent communication, and whether the core offers training and inference as a service
WT#4_SensingKI#20Integrated Sensing and Communication — sensing function discovery, configuration, data collection and exposure of results
WT#5_Data_FrameworkKI#216G Data Framework in SA2 — discovery, collection, transfer, labelling, storage, plus consent and governance
WT#6_ComputeKI#226G Computing Support — compute across handset, core and application server; authorisation and site reselection

Three of those pairings are stated: WT#3 owns KI#18 and KI#19, and WT#5 owns KI#21. That WT#4 maps to KI#20 and WT#6 to KI#22 is this newsletter’s inference from matching names, and remains so.

The line SA2 drew through the middle of AI

The split between KI#18 and KI#19 is the most consequential structural decision in the beyond-connectivity half, and it is a familiar one under new numbers. Zahid Ghadialy, writing at free6gtraining on 8 August 2026, gives the cleanest statement of it: AI for network is “using AI to improve the operation of the communications network itself”; network for AI asks “how the communications network should evolve to support distributed AI workloads, AI agents, model training and inference”.

One is a network that thinks; the other carries other people’s thinking. Separating them lets the harder political question — whose agents get to talk through the operator’s core, on what terms — sit in its own key issue instead of contaminating the rest.

“3GPP has previously made clear that it does not currently intend to specify particular AI/ML models.”

That line, also Ghadialy’s, is the boundary the whole task works inside. SA2 is specifying where models are invoked, who authorises them and how results move — not what is in them.

Twenty-eight files, filed the week SA2 met

The drafting folder for KI#18 in SA2’s open working area holds twenty-eight files, all uploaded between 18 and 27 August 2026 — a window that straddles SA2#176, Prague, 24–28 August. Every edition since that round has reported no public outcome from it. This is still not an outcome — but it is the first sight of the work.

No file here was opened; Word documents on the 3GPP server return 403 to this newsletter, as they have all series. What the listing gives is titles, dates, sizes and revision chains — and the titles alone are worth reading.

DraftTitle, as filed
S2-2608455Solution updates for “Transport of intent from UE and AF” — revised six times in little over an hour on the evening of 18 August
S2-2608459“Resolving EN for requests without intent”
S2-2608464“Capability exchange and NW-Agent based procedure for request-based monitoring”
S2-2608861“Tool invocation and multi-agents communication”
S2-2608853“New separated domain from PS domain, within 6G CN”
S2-2608875“An Enhanced NEF capable of handling Intent based closed loop operations” — the one file naming two key issues, 18 and 21

Three things fall out. The recurring noun is intent — an abstract request the network is expected to decompose rather than a parameter it is told to set. Tool invocation is vocabulary borrowed wholesale from the agent frameworks of the last two years, now sitting in a 3GPP drafting folder. And the solution variants are numbered as high as #18.22. Whatever else KI#18 is short of, it is not proposals.

Tingyu Xin of Samsung R&D Institute UK and Hyesung Kim of Samsung Research describe the same work from inside: a core network that processes “intent-based requests” and “dynamically compose[s] procedures” to fulfil them, with multi-vendor interoperability as a constraint rather than an afterthought.

The Nokia contribution Edition 14 quoted — filed to SA2#174 in Malta, April 2026 — proposed confining the agent discussion to KI#18 and KI#19, other key issues documenting functionality “i.e. not referring to AI agents explicitly but not excluding it either”. One company’s proposal, not an agreement. Four months on, the AI folder is the busiest in the tree — consistent with the fence having been built.

What a file listing proves, and what it does not

This edition rests on a directory listing. That is a real primary source and a narrow one, so here is exactly what it does and does not support.

Not one of the twenty-eight documents was opened. Every title, date and size above is the file name as the server prints it. Three of the twenty-eight are 0 KB. That a draft exists says nothing about whether it was discussed, agreed, or withdrawn — and #18.22 is the highest variant number visible, not a count of solutions.

The meeting attribution is ours. The timestamps run 18–27 August and SA2#176 ran 24–28 August in Prague; the folder is not itself labelled with a meeting. Read it as activity in that window, not as minutes.

Only KI#18 opened. The folders for KI#19 through KI#22 were each requested and refused, so nothing here implies anything about their volume. A path of the form Pen-holder Drafts/KI#22 appears in search results and could not be opened; it is printed because it exists, not because it was read.

And the five key-issue objectives come from v0.4.0 of the report, 27 February 2026, through a third-party mirror, where the portal is at v0.8.0 of 28 July. Key-issue scope is among the more stable parts of an SA2 study; the solutions are not, which is why none is quoted.

Sensing, data and compute get what is left

The other three tasks are not smaller problems. KI#20 has to make a base station a sensing instrument and expose the results only to whoever is entitled to them. KI#21 has to invent a data framework for a network that generates training data as a by-product, and it carries the consent and governance question explicitly. KI#22 has to let compute be discovered, authorised and reselected across handset, core and cloud.

The constraint they share is the one Edition 14 counted. FS_6G_ARC closes on 3 March 2027, and after Prague only three SA2 meetings remain, the last ending five days before the deadline. Three meetings, twenty-four key issues — and the one whose solution variants run to #18.22 is not the one that gets dropped.

Where it goes next

The chain is already visible downstream. Abdessamad El Moatamid and Xuefei Zhang of Huawei and Zhenning Huang of China Mobile, all CT WG3 delegates, set it out on 3gpp.org: SA1 requirements in TR 22.870 and TS 22.270, SA2 architecture in TR 23.801-01, then CT3’s Stage 3 study FS_6G_AIP_CT, which is weighing four existing open protocols — A2A, ANP, AGNTCY and MCP — rather than starting from nothing. Whatever KI#18 and KI#19 settle becomes CT3’s input.

Sources

Daily Dose of 6G — tracking 3GPP Release 20's 6G study phase, one study at a time. Timeline after 3GPP's published Release 21 schedule.