KanDDDinsky agent-swarm workshop results
Итоги KanDDDinsky: воркшоп с роем агентов
C · Weak signal Short Telegram post by Rinat Abdullin (LLM под капотом) reporting activity counts from a two-day conference workshop; first-hand and the memory contents are published, but there is no outcome measure or analysis in the post.
Summary based on the full text·Original: about 2 minutes to read
In short
A Telegram post (17 October 2026) on the channel LLM под капотом ("LLM under the hood") by Rinat Abdullin, summarising workshops at the KanDDDinsky conference. He put the memory layer of his "Personal OS" on a server, let participants connect their own subscription agents, and added two tools: publishing slides and running Python in a sandbox to repair a broken experiment (BitGN ECOM1 tasks). Participants steered the swarm by filling the shared memory and instructions.
Key numbers
- 20workshop participants whose own ChatGPT and Claude agents shared one memory layer
- 8,650sandbox runs over two days
- 697changes logged over the two days
- 150experiments, producing 1.4 GB of data
As reported by the source.
Russian·Web page·opens t.me
What it reports
- Over two days and 20 participants: 8,650 sandbox runs, 697 changes and 150 experiments with 1.4 GB of data.
- Almost all experiments came from several agents working together.
- Participants and their agents invented a messaging system, tried voting and sabotage, and filled
AGENTS.mdwith junk. - Progressive disclosure is built into the memory: "docs" is the root folder and a
treeinstruction is in the system's start-up instructions, read beforeAGENTS.md. - Slides, usage statistics and the memory contents are published and left in place.
How they got it
Activity logs from the shared memory server during two workshops. No measure of whether the experiment was fixed, or how well, is given in the post.
Limits
- Counts measure activity, not results; the post gives no benchmark or quality figure.
- A conference workshop with self-selected participants, run by the tool's author.
- The linked statistics and slides were not reviewed for this digest.
Why it matters for companies
The interesting signal is behavioural: when many people's agents share one memory, the shared instruction file degrades fast unless someone owns it. That matters for any team planning a common AGENTS.md or knowledge layer for agents. As evidence of productivity it is too thin to cite.
Key passages in translation
Russian original
мы будем управлять этим роем заполняя память и инструкции для них.
We will steer this swarm by filling in the memory and instructions for them.
Russian original
Причем практически все эксперименты были результатом совместной работы нескольких агентов.
And practically all the experiments were the result of several agents working together.
Russian original
придумали систему сообщений, попытались сделать голосование и саботаж, и забили AGENTS.md хламом.
They invented a messaging system, tried to set up voting and sabotage, and stuffed AGENTS.md with junk.
Summary based on the full text. Read the original for the full argument and data.Read the original on t.me
- Tags
- AI enablement, Multi agent