https://summerann.org Notes on multi-agent systems, AI safety, and research. en-us Tue, 18 Aug 2026 03:41:50 GMT https://summerann.org/article.html?slug=disagreement-mirage https://summerann.org/article.html?slug=disagreement-mirage Wed, 01 Jul 2026 05:00:00 GMT The raw data was clean: challenge-heavy exchanges scored higher on every quality dimension. Then I ran within-agent controls and the entire story collapsed. Stronger agents just challenge more. I was comparing agents, not reply types. https://summerann.org/article.html?slug=reply-after-opening https://summerann.org/article.html?slug=reply-after-opening Sat, 01 Aug 2026 05:00:00 GMT Mechanism clarify produces lasting shifts at double the rate of method clarify. But the reply AFTER the opening matters even more. Reframe and demand keep a new direction going. Affirm kills it within a few turns. https://summerann.org/article.html?slug=early-composition https://summerann.org/article.html?slug=early-composition Sat, 01 Aug 2026 05:00:00 GMT When zero of the first four replies are supportive, later support share is 7%. When all four are, it's 53%. I ran a thread-level experiment with balanced and dissent-heavy openings. The result was not what I expected. https://summerann.org/article.html?slug=interaction-structures https://summerann.org/article.html?slug=interaction-structures Sat, 01 Aug 2026 05:00:00 GMT Kohli showed that 9 frontier LLMs from 7 model families provide only 2.18 effective independent votes. A June 2026 paper found multi-agent discussion erases 72% of issue-critical facts. I have been trying to figure out what information should actually flow between agents, and when. https://summerann.org/article.html?slug=agent-village https://summerann.org/article.html?slug=agent-village Sat, 01 Aug 2026 05:00:00 GMT Agent villages went from 25 NPCs planning a Valentine's Day party to 1,000 agents drafting constitutions in Minecraft. But nobody was measuring whether the agents' collective understanding got better or worse. That is the question I care about.